> ## Documentation Index
> Fetch the complete documentation index at: https://docs.asteragents.com/llms.txt
> Use this file to discover all available pages before exploring further.
# Changelog
> Stay up to date with the latest releases from Aster Agents
### Claude Opus 5
Anthropic's new **Claude Opus 5** is now available to select for your agents.
Pick it under **Control Hub → Agents → Model** (or your org's default-model
setting) as `Anthropic: claude-opus-5`.
* **Top-tier reasoning:** the newest Opus-tier model, priced at $5 / input MTok
and $25 / output MTok — the same rate as Opus 4.8.
* **Drop-in:** adaptive thinking and a 1M-token context window, so existing
agents work without changes.
Existing agents keep their current model — switch to Opus 5 whenever you're ready.
### Paycom Integration
Connect **Paycom** and let your agents read your HR and payroll data — employee
directory, timecards and punches, positions, labor allocations, and new hires.
Read-only, so agents can report on your workforce without ever changing anything
in Paycom.
* **Query the workforce:** ask for active headcount by department, recent hires,
hours by employee, labor-allocation breakdowns, and more — agents pull it live
and return structured results.
* **Built for KPIs:** large pulls stream the full result set to a CSV the agent
analyzes in code, so directory-sized queries don't crowd the conversation.
* **Sensitive data stays off by default:** rate, salary, tax, and SSN endpoints
are blocked unless you explicitly enable them — with Paycom, and in the agent.
**Before you connect:** Paycom only accepts API calls from allow-listed IP
addresses, so your Paycom administrator adds Aster's static IPs to your API
user's allow-list first. Once that's done, set it up in
**Control Hub → Integrations → Paycom** with your API SID and Token — the data
center is detected automatically.
### Admin Portal
If you administer more than one Aster organization, there's now a single place
to see and manage them all: the **Admin Portal** at
[asteragents.com/admin](https://www.asteragents.com/admin).
* **Every org at a glance:** conversations, active users, members, agents, and
knowledge bases for each org you admin, with search and sortable columns.
Expand a row for estimated AI spend by model — the same cost math as each
org's own dashboard.
* **Copy across orgs:** move agents, knowledge bases, and skills from one org
into another — or into a brand-new org created on the spot. Dependencies come
along automatically (an agent's knowledge bases and skills), API keys only
move when you explicitly say so, and a **preview** shows exactly what will be
copied before anything is written.
* **Invite members anywhere:** add people to any org you admin without
switching organizations — existing users are added instantly, new emails
receive an invitation.
* **Stand up a workspace in one step:** the copy dialog can create a new org,
seed it with agents, and invite its members all in a single action.
Available to any organization admin — sign in and visit **/admin**.
### NetSuite (SuiteTalk) Integration
Connect a **NetSuite** account and let your agents work with your financials
and transactions directly. Authentication uses NetSuite's Token-Based Auth,
so no passwords are stored.
* **Query with SuiteQL:** ask questions across records — invoices, transactions,
customers, and more — and get structured results back.
* **Read records:** fetch a single record by type and id, or list records of a
type with filters.
* **Write when you grant it:** create, update, and delete records — an optional,
separately granted tool, so read-only stays read-only.
* **Multiple accounts:** connect more than one NetSuite instance — add another
from Control Hub with a different label.
Set it up in **Control Hub → Integrations → NetSuite (SuiteTalk)**.
### Value & ROI
You can now put a dollar value on what an agent's work is worth — and see it
measured against the cost the platform already tracks for every run. The result
is **ROI**, per conversation and rolled up across your team.
* **Set a value rule** on any agent's config page (the new **Value** card):
either a flat amount for every conversation it has, or an amount tied to an
outcome — like **\$50 each time it sends an email** — using a tool the agent
already has.
* **See it in the Inbox:** each conversation shows the value it earned next to
its cost.
* **Roll it up on the dashboard:** the **Analytics** tab gains an **ROI** stat,
plus per-user and per-agent value in the existing breakdowns.
* **Entirely optional:** leave it off and nothing changes. Before an agent earns
anything, ROI reads **N/A** — never a misleading `$0`.
Value is an estimate you define (a proxy for the outcome an agent drives), not
booked revenue. See [Build an Agent → Value & ROI](/features/build-an-agent#value--roi-optional).
### Retell AI Integration
Connect **Retell AI** and let your agents drive voice calls end to end — place
outbound calls, read back transcripts and post-call analysis, and manage the
Retell voice agents and LLM configs behind them.
* **Place calls:** an agent can dial a number with a personalized script
(dynamic variables) and tag the call with metadata.
* **Pull results:** retrieve a call's transcript, summary, sentiment, success,
and any extracted data — or list and filter recent calls.
* **Manage config:** list, get, create, and update Retell agents and LLMs; list
phone numbers and voices.
* **React to completed calls:** an optional signed webhook hands a finished
call's transcript and analysis to an Aster agent so it can follow up
automatically.
Connect it under **Control Hub → Integrations → Retell AI** with your Retell API
key, then enable the **Retell** tools on any agent. See the
[Retell AI integration guide](/integrations/retell) for setup.
### Favorite Apps
You can now **star the apps you use most** so they're always easy to find. Each
person's favorites are their own — starring an app is personal and invisible to
the rest of your team.
* **Star from anywhere:** hover any card on the **Apps** grid and click the star,
or use the star in an app's header when it's open.
* **Favorites float to the top** of the Apps grid in every sort order, so your
go-to apps are always the first thing you see.
* **Filter to just favorites** with the **★ Favorites** toggle next to the search
box.
### Build Agents on xAI's Grok 4.5
You can now pick **Grok 4.5** as the model for any agent. It's xAI's most
capable model yet — built for coding, agentic tasks, and knowledge work —
with a **500K-token context window** and image understanding.
* **Select it anywhere you choose a model:** it appears under **xAI** in the
model picker for agents and chat.
* **High-reasoning by default** with a large context window, so it handles
long documents and multi-step tool use well.
Connect your **xAI** API key under **Control Hub → Integrations** to enable it.
### Fireworks AI Provider
You can now connect **Fireworks AI** and run open-source models in your agents.
Add your API key in **Control Hub > Model Providers > Fireworks AI**, then pick a
Fireworks model from the model picker in **Control Hub > Edit Agent**.
Two models are available to start:
* **GLM 5.2** (`fireworks:accounts/fireworks/models/glm-5p2`) — strong all-round
open model with a 1M context window. $1.40 / $4.40 per 1M input / output tokens.
* **Kimi K2.6** (`fireworks:accounts/fireworks/models/kimi-k2p6`) — 262K context
with vision support for chat-attached images. $0.95 / $4.00 per 1M tokens.
Both run on Fireworks' serverless tier — pay per token, no cold starts. More
open-source models can be added on request.
### GPT-5.6 (Sol, Terra, Luna)
OpenAI's **GPT-5.6** family is now available in Aster Agents, in three tiers so
you can match capability to cost:
* **GPT-5.6 Sol** (`openai:gpt-5.6-sol`) — frontier model for complex reasoning
and coding. Runs with **high** reasoning effort. $5 / $30 per 1M input / output tokens.
* **GPT-5.6 Terra** (`openai:gpt-5.6-terra`) — balances intelligence and cost.
$2.50 / $15 per 1M tokens.
* **GPT-5.6 Luna** (`openai:gpt-5.6-luna`) — optimized for cost-sensitive,
high-volume workloads. $1 / $6 per 1M tokens.
All three support a 1M context window, 128K max output, vision, and reasoning
tokens. Select one from the model picker in **Control Hub > Edit Agent**. New
OpenAI agents onboarded from today default to GPT-5.6 Sol.
### Aster Is Now an MCP Server
You can connect **Claude Desktop, Claude Code, the Claude web app, Cursor, and
other MCP clients** directly to Aster. Once connected, the client can list and
invoke your agents, search your knowledge bases, and read your conversations —
all as you, scoped to your organization.
* **One URL:** add `https://www.asteragents.com/mcp-server` as a custom
connector. Sign-in uses your normal Aster login.
* **Run your agents from anywhere:** `call_agent` invokes an agent and the run
shows up in your dashboard tagged `mcp`, alongside chat and email.
* **Build through your agents:** create or edit agents and apps by asking a
builder agent in your org — no separate tooling needed.
See [Aster MCP Server](/features/mcp-server) to connect a client.
### Ask Questions Right Inside an App
Apps can now include a **live chat panel**, so anyone viewing a dashboard or
report can ask the agent a question without leaving the page. The reply streams
in right there — same agent, same tools, same charts, tables, and citations as
the main chat.
* **Ask for it in plain language:** *"add a chat panel to this app so people can
ask questions about the data."* An **Ask about this app** panel appears.
* **It talks to the app's own agent** (or one you name), so it already knows the
subject matter and can use its tools to pull fresh detail on demand.
* **Great for live demos and diligence** — drill into a number or ask *"which
companies are at risk and why,"* inline while looking at the dashboard.
Each visit starts a fresh conversation. See [Apps](/features/apps) to try it.
### Read Your Cobalt Portfolio Data with Agents
You can now connect **Cobalt** (a FactSet company) and let your agents read your
private-equity portfolio-monitoring data — **read-only**, so nothing in Cobalt
ever changes.
* **One tool, the whole API.** Enable **Cobalt: Read Portfolio Data** and your
agent can reach portfolios, funds, deals, portfolio companies, cashflows,
valuations (IRR, MOIC, DPI, TVPI, NAV), and operating metrics.
* **It finds its own way around.** The tool is self-describing — the agent looks
up the available endpoints and fields on its own, then pulls exactly what your
question needs.
* **Big pulls stay fast.** Large result sets spill to a downloadable file the
agent can crunch in Python, so a fund-wide export never floods the chat.
An admin connects it once for the whole org under **Control Hub → Integrations →
Cobalt** with your Cobalt API key. See the [Cobalt integration guide](/integrations/cobalt)
to get started.
### Feedback Message: Fill in the User's Details Automatically
The custom message on the conversation feedback dialog now supports **template
variables** that fill in the current user's details the moment it renders — so
a link can carry that user's context straight into your own systems.
* **`{{user.email}}` and `{{user.name}}`** — the signed-in user's identity.
* **`{{attr.KEY}}`** — any custom attribute your organization stores on a user.
For example, if you keep an internal ID on each user, `{{attr.vnum}}` drops it
right in.
The most common use is deep-linking to your own support portal with the user
already identified — no login lookup on your side:
```
https://portal.example.com/support?email={{user.email}}&id={{attr.vnum}}
```
A few details: unknown variables render as empty (never a stray placeholder),
attribute keys are case-sensitive, and the message is still sanitized before it
renders. Identity (`user`) and your custom attributes (`attr`) live in separate
namespaces, so a custom attribute can never override the real signed-in email.
Set it in **Control Hub → Settings → Feedback Modal**.
### Popularity Comes to the Apps Page
Finding the right app in an org with dozens of them just got easier. The Apps
page now shows how much each app is actually being used — and puts the popular
ones first.
* **A momentum sparkline on every tile** — a small 14-day trend of opens, with
the total, right on the card. It runs **green when an app is picking up
steam** and **red when it's cooling off**, so you can tell at a glance what
your team is leaning on lately.
* **"Most used" is the new default sort** — the grid opens with the apps people
actually rely on at the top, without you touching a thing. Prefer the old
order? "Recently updated," "Name," and "Agent" are still in the sort menu.
* **It fills in on its own** — usage starts counting now; the more an app gets
opened, the richer its trend line becomes.
Head to **Apps** and the busiest ones are already waiting at the top.
### Download Your Apps as PDF or PNG
Every Aster App now has an **Export** button, right next to "Ask about this
app." Turn a live dashboard into a shareable document — no need to ask the
agent to build export logic into the app itself.
* **PDF** — the view you're looking at, paginated. Charts, gradient cards,
and tables render exactly as they appear on screen (full color, none of
the browser's print header/footer), and pages break *between* sections so
a chart, card, or table row is never cut in half across a page.
* **PNG** — the same view as a single high-resolution image, ideal for
dropping into a deck or an email.
* **Landscape or portrait** — your choice, from the same menu.
It's fully client-side: nothing about your data leaves the browser to
generate the file.
**Sharper page breaks, when you want them.** Agents building apps can add
standard CSS break utilities — `break-inside-avoid` to keep a card or table
whole, `break-before-page` to start a section on a fresh page — and the
export honors them. Existing apps export as-is; the pagination just gets
tighter when an app opts in.
Open any app, hit **Export**, and pick PDF or PNG.
### Upload Many Files at Once — Reliably
Sending an agent a big batch of attachments in a single message is now
solid. Previously, uploading many files (or very large ones) at once could
exceed the model's context window and cause the turn to come back empty.
That's fixed.
* **Batches that fit are used inline; the rest stay one tool call away.**
When an upload is too large to place entirely in context, Aster keeps
every file attached and lets the agent pull what it needs on demand with
**Read File** or **Execute Python** — nothing is silently dropped.
* **A wider range of files won't break a message.** An unsupported image
type (such as SVG) mixed into an upload no longer fails the whole turn —
it's simply made available to read like any other attachment.
Drop in 5 files or 50, of mixed types, and the agent responds and works
through them.
### Claude Sonnet 5 Is Live on Aster
Anthropic's newest Sonnet, **Claude Sonnet 5**, is available on Aster today — and it's the model we'd reach for first on a new agent. It hits near-Opus quality on coding, tool use, and agentic work, but costs meaningfully less than Opus, so it's the latest and greatest *and* the value pick for most agents.
**What's new:**
* **Available everywhere** — select `claude-sonnet-5` in the model picker for any agent, or set it as your org's default Anthropic model.
* **Latest and greatest, cheaper than Opus** — `$3` per million input tokens and `$15` per million output, versus Opus 4.8's `$5` / `$25`. Anthropic is running introductory pricing of `$2` / `$10` per million through 2026-08-31.
* **1M context and adaptive thinking** — the same 1M-token context window and adaptive extended thinking as the rest of the current Claude lineup, plus high-resolution vision.
Open any agent, pick **claude-sonnet-5** from the model dropdown, and give it a run.
### Inbound Emails Now Carry Sender, Recipients, and Subject
When you email an agent, it now sees the full envelope — not just the body. Each inbound message is prefixed with the sender, the visible recipients, the subject, and whether the agent was BCC'd:
* **From / To / Cc** — the agent can now read who sent the message and who it was addressed to, using the visible header recipients (so an agent's own address isn't shown as noise).
* **Subject** — included alongside the body.
* **BCC awareness** — if the agent was BCC'd, the message says so explicitly. BCC'd mail is still ingested silently and never triggers a reply, but now the agent *knows* it was BCC'd and can act on that.
This unlocks recipient-based routing: agents can branch on who's on the thread or whether they were BCC'd — for example, silently logging an interaction when BCC'd versus running a confirmation step when addressed directly.
### Every New Agent Starts with Read File and Execute Python
New agents now begin with two safe, broadly useful tools already enabled, so they start with a sensible baseline instead of an empty toolbox:
* **Read File** — open a single file the agent can already reach (a sandbox file, a conversation attachment, a knowledge-base file, or a URL). Documents auto-extract to text; images come back as vision; large files are line-paginated. It's a brand-new tool reference — see [Read File](/tools/read_file).
* **Execute Python** — run code in a sandbox. It pairs naturally with attachments and knowledge-base files for parsing and transforming uploaded data.
**Default-on, not forced.** Both behave like any other tool — uncheck either one on any agent and it stays off. The defaults are applied wherever an agent is created: the Control Hub form, the API, and agents built by the **Build Agents** agent via `manage_agents`.
**Also:** images returned from tools now render on Google (Gemini) models, joining Anthropic, OpenAI, and Azure. On xAI models, image reads gracefully fall back to text.
### Agent Version History: Every Save Is Recoverable
Every time you save an agent, its configuration — system prompt, model, tools, and connected knowledge bases, skills, and MCP servers — is now captured as a version automatically. No extra step, nothing to turn on.
Open any agent and click **History** to see the full timeline of changes, newest first, with the version currently in use marked **Current**. Made an edit you want to undo? Click **Restore** on any earlier version to roll the agent's configuration back to that point — it takes effect immediately.
* **Automatic:** versions are recorded on every save, including changes made by tools, skill updates, and tool-schema syncs — not just manual edits.
* **Safe restore:** restore replaces the current configuration, so it's blocked while you have unsaved edits — save or discard them first.
* **No clutter:** identical saves are de-duplicated, so history only grows when something actually changes.
### Org Secrets: Store Credentials Your Agents Can Use, Never See
Some tasks need a real login — signing into a vendor portal to pull invoices, calling an API that isn't a built-in integration. Org admins can now store those credentials once and let agents use them in any tool call **without the secret ever being exposed**.
Add secrets in **Control Hub > Integrations**, at the bottom of the page below MCP Servers. Each secret is a name and a value (for example, a portal username and password). An agent references one in a tool call with a `{{secret:KEY}}` placeholder — for instance, having **Browser Use** sign into a site with `{{secret:SITEONE_USERNAME}}` and `{{secret:SITEONE_PASSWORD}}`.
**Built so the value never leaks:**
* The real value is injected server-side at the moment the tool runs, then **redacted from the tool's output** — so the model only ever sees the placeholder and a masked result, never the plaintext.
* Secrets are **write-only**: once saved, a value can be rotated or deleted but never viewed again — not in the UI, not over the API, not even by an admin.
* Management is **admin-only**, and stored values are never returned to the browser for anyone in the org.
Copy a secret's `{{secret:KEY}}` reference straight from the list to drop into an agent's prompt or tool configuration.
### Apollo.io: Find and Enrich B2B Contacts
Agents can now tap [Apollo.io](/integrations/apollo) for B2B contact and company data — verified work emails and firmographics from one of the largest prospecting databases. Two tools that chain together:
* **[Apollo Prospect](/tools/apollo_prospect)** — discover *new* people or companies matching an ideal customer profile (job title, seniority, location, company domain, employee count, revenue, keywords). Search is free and returns each match's `id`.
* **[Apollo Enrich](/tools/apollo_enrich)** — fill in *known* people (by Apollo `id`, email, LinkedIn URL, or name + company/domain) or companies (by domain) with verified work emails and firmographics.
**Credit-safe by design:** prospecting search spends no credits, so you can build and qualify a list for free, then spend roughly one export credit per contact to reveal the verified email. The reliable workflow is prospect → take the `id` → enrich.
Connect with your Apollo API key in **Control Hub > Integrations**. See the [Apollo.io integration guide](/integrations/apollo) to get started.
### Add a Custom Message to the Feedback Dialog
Org admins can now add a custom message that appears in the conversation feedback dialog — shown whenever a user rates a conversation, whether they mark it helpful or not. Use it to point people to your customer portal, support team, or help center when chat alone didn't get them all the way there.
Set it in **Control Hub > Settings > Feedback Modal**. Basic HTML is supported — links, paragraphs, and bold — so you can drop in a short line like "Still stuck? Talk to our team" that links out to your help center or support portal.
**Safe by design:** the message is sanitized before it renders, so scripts and event handlers are stripped. The setting is admin-only and empty by default, so nothing changes for your users until you configure it.
### MCP Connections Now Stay Alive — and Tell You When They Don't
OAuth connections to [MCP servers](/integrations/mcp-servers) now refresh their tokens automatically in the background, including against providers that rotate refresh tokens on every use — so a connection made once keeps working indefinitely on a healthy server.
And when a connection genuinely can't be recovered (revoked or expired on the provider's side), Aster now says so instead of failing silently:
* The server card in **Control Hub > Integrations** shows a **Connection error** badge with a one-click **Connect** button to reauthorize.
* Chat shows a banner above the composer — *"Your connection needs to be reconnected before this agent can use its tools"* — the moment the problem is detected.
Previously an expired connection still displayed as Connected while agents quietly lost access to the server's tools. Transient provider outages don't trigger the prompt — those retry automatically.
### Sage Intacct: Query Your Accounting Data
Agents can now read directly from [Sage Intacct](/integrations/sage-intacct) — GL detail, AP bills, AR invoices, vendors, customers, the chart of accounts, and every other object the connected user can see. Two tools that chain together:
* **Sage Intacct Get Schema** (`intacct_get_schema`) — lists the objects available in your Intacct company and returns exact field names, types, and relationships for any of them, so agents discover the data model at runtime instead of guessing.
* **Sage Intacct Query** (`intacct_query`) — reads records from any object with a SQL-style filter predicate (e.g. `TOTALDUE > 1000 AND STATE = 'posted'`). Auto-paginates up to 50,000 rows, and **large result sets attach as a CSV file** — analyze them with [Run Code](/tools/execute_python) or deliver as a download — while the in-conversation payload stays small.
**Read-only by design:** the tools only issue Intacct read functions, so agents can explore the books freely without any risk of modifying the accounting system. Permissions follow the connected Web Services user, so scope its roles to control what agents can see.
Connect with your Sender ID and a Web Services user in **Control Hub > Integrations**. See the [Sage Intacct integration guide](/integrations/sage-intacct) to get started.
### Claude Fable 5 Is Now Available
Anthropic's most capable model, **Claude Fable 5**, is now selectable for any agent. It pairs a **1M-token context window** with up to **128k output tokens** and vision support — built for the most demanding reasoning and long-horizon agentic work.
Like the Opus 4.x models, Fable 5 runs with **adaptive thinking** always on, so it scales its reasoning depth to the task automatically.
Pricing is **$10 per million input tokens** and **$50 per million output tokens**. Existing agents keep their current model — choose Claude Fable 5 from the model picker in **Control Hub** (requires a connected Anthropic API key).
### Lusha: Find and Enrich B2B Contacts
Agents can now tap [Lusha](/integrations/lusha) for B2B contact and company data — verified emails, direct phone numbers, and firmographics. Two tools that chain together:
* **[Lusha Prospect](/tools/lusha_prospect)** — discover *new* people or companies matching an ideal customer profile (department, seniority, location, company size, revenue, industry, technologies, intent). Returns non-PII previews plus, for each match, which data points are available to reveal.
* **[Lusha Enrich](/tools/lusha_enrich)** — fill in *known* contacts or companies (by name + company, email, LinkedIn URL, or domain) with emails, phones, and firmographics.
**Credit-safe by design:** reveals are billed per data point (email = 1 credit, phone = 5), so Lusha Enrich reveals emails by default and leaves phones off unless you ask for them. Prospecting and preview lookups never reveal PII — build and qualify a list cheaply, then spend credits only on the contacts worth reaching.
Connect your Lusha API key (Premium or Scale plan) in **Control Hub > Integrations**. See the [Lusha integration guide](/integrations/lusha) to get started.
### Apps: Agents Can Now Browse and Restore Version History
Every update to an [Aster App](/features/apps) has always created a preserved version behind the scenes — now agents can see them. The **manage\_apps** tool gains two actions:
* **`list_versions`** — returns an app's publish history: when each version was published, which is current, and what source files it contains.
* **`get_version`** — returns the complete source code of any historical version.
And `load_to_sandbox` now accepts a `version_id` — loading a historical version's source **and its snapshot data and assets** into the sandbox, so `publish_from_sandbox` can restore it with full fidelity (essential for data-driven apps whose versions differ only in data).
Together they let an agent answer "what changed in this app?" and recover from a bad update — restore an older version with no data loss, URL unchanged. Works retroactively for every existing app.
See the [Manage Apps tool docs](/tools/manage_apps) for details.
### HubSpot Search: Automatic Pagination & Bulk CSV Export
The **hubspot\_search** tool now handles large result sets the same way [Salesforce queries](/integrations/salesforce) do — ask for the data, get a CSV.
**What changed:**
* **Automatic pagination** — agents set `limit` to the total number of records they want (up to 10,000, HubSpot's cap per search) and the tool pages through results internally, including rate-limit pacing and retries. Manual cursor pagination is gone.
* **CSV file attachments** — when results are too large to fit in the conversation, the full row set is attached as a CSV file. Agents can analyze it with [Run Code](/tools/execute_python) or deliver it as a download.
* **Beyond 10,000 records** — the tool guides agents to partition searches with filters (e.g., date ranges) when a query matches more than HubSpot's per-search cap.
Bulk pulls like "export all contacts with their lifecycle stage" are now a single tool call instead of dozens of paginated requests.
### Connect MCP Servers with OAuth, Not Just API Keys
Remote MCP servers can now authenticate with **OAuth**. Instead of sharing one API key across your whole org, each person connects their **own account** to a server, and their agents call its tools with their own token.
**What's new:**
* **Two ways to authenticate** — choose **API Key** (a shared header like a Bearer token, sent on every request) or **OAuth** (each user signs in individually).
* **Per-user connections** — an admin adds a server once at the org level, then each teammate clicks **Connect** and authorizes with their own account. Tools run with that user's permissions.
* **Fits how providers issue OAuth** — supports both **Dynamic Client Registration** (for servers that support it) and **pre-created OAuth apps** (paste a client ID and secret, and whitelist Aster's callback URL).
* **Status at a glance** — connect, reconnect, or disconnect from **Control Hub > Integrations**, with an indicator on each agent showing whether you've connected the servers it uses.
**Getting started:**
1. Go to **Control Hub > Integrations** and scroll to **MCP Servers**
2. Click **Add Server**, choose **OAuth**, and paste the server's URL
3. Click **Connect**, sign in, then assign the server to any agent
See [MCP Servers](/integrations/mcp-servers) for the full guide.
### Apps Can Now Be Locked to Their Author
Once an agent builds an app you're happy with, you can **lock** it so it doesn't get changed out from under you. While an app is locked, only the person who created it — or an organization admin — can edit, delete, or unlock it. Everyone else can still open and view it as normal.
**What you can do:**
* Tell any agent with the **Manage Apps** tool to *"lock this app"* — only you (the author) or an admin can change it afterward
* Tell it to *"unlock the app"* whenever you want to allow edits again
* See a **Locked** badge on locked apps in the **Apps** list and Control Hub
* Apps are **unlocked by default** — nothing changes for your existing apps until you choose to lock one
If another person's agent tries to edit a locked app, it gets a clear message explaining the app is locked and won't overwrite your work. This is handy when an app becomes a shared source of truth — a board report, a portfolio dashboard, a finished deliverable — and you want it frozen.
See [Apps](/features/apps#locking-apps) for details.
### Claude Opus 4.8 Is Live on Aster
Anthropic's newest and most capable model, **Claude Opus 4.8**, is available on Aster today. Pick it from the model dropdown on any agent — it joins the existing Claude lineup with the same 1M-token context window and adaptive extended thinking.
**What's new:**
* **Available everywhere** — select `claude-opus-4-8` in the model picker for any agent, or set it as your org's default Anthropic model.
* **New Anthropic default** — newly created Anthropic agents now start on Opus 4.8 out of the box. Existing agents keep whatever model they're already on; switch them over anytime from the model dropdown.
* **Same pricing as Opus 4.7** — `$5` per million input tokens and `$25` per million output tokens, unchanged. No cost surprises in your usage rollups.
Open any agent, pick **claude-opus-4-8** from the model dropdown, and you're on the latest.
### Apps: Your Agents Can Now Build You Live Dashboards
When you ask an agent to analyze data, the results used to live and die in the chat thread. Now your agents can create **persistent, interactive apps** — dashboards, reports, and data views that live at a permanent URL inside Aster. Bookmark them, share them with your team, and come back whenever you need them.
**What you can do:**
* Ask any agent to *"build me a dashboard"* from your GA4 traffic, Salesforce pipeline, Snowflake data, QuickBooks financials, or any other connected source
* Get back a live, interactive page with charts, tables, KPI tiles, tabs, and more — not a screenshot, not a PDF
* Share the link with anyone in your org — they see the same view with the same data
* Tell the agent to *"update the dashboard"* whenever you want fresh numbers
* Pair with [Scheduled Tasks](/features/scheduled-tasks) to auto-refresh daily or weekly
**Getting started:**
1. Add the **Manage Apps** tool to any agent
2. Ask the agent to build what you need
3. Open the `/apps/{id}` link it gives you
Browse all your org's apps from the **Apps** link in the sidebar.
See [Apps](/features/apps) for the full guide.
### Compute Route — Travel Time and Distance, No Setup Required
Every agent on Aster now has a new built-in tool: **Compute Route**. Give it a starting point and a destination — or hundreds of them — and it returns travel time and distance via Google's Routes API. Driving, walking, bicycling, and transit are all supported, with live-traffic-aware driving by default.
Unlike the Google Search Console and Google Analytics integrations, this is a **platform tool, not an integration**. No OAuth, no per-org connection, no API key for admins to provision — Aster covers the Google Maps Platform usage at the bundled price.
**What's included:**
* **Flexible waypoints** — pass an address (`"1 Hacker Way, Menlo Park, CA"`), a place name (`"Eiffel Tower"`), or a raw `"lat,lng"` string. The Routes API geocodes internally, so no separate lookup step.
* **Two bulk modes:**
* **Matrix mode** — provide `origins[]` and `destinations[]` and get every pair back (N × M). Best for all-pairs comparisons like "of these 8 warehouses, which is fastest to each of these 12 customers?". Capped at 200 element pairs per call.
* **Paired mode** — provide `pairs: [{from, to}, ...]` and get N independent A→B results back. Best for "drive time for these 50 driver/customer pairs in my fleet." Dramatically cheaper than matrix mode for the same number of results — 50 paired lookups bill \~50 elements versus 2,500 for a 50×50 matrix. Capped at 100 pairs per call.
* The two modes are mutually exclusive; the tool errors with a clear message if both are passed.
* **Travel modes and traffic** — `driving` (default), `walking`, `bicycling`, `transit`. Driving accepts an optional `departure_time` (RFC3339) to plan against a specific future commute, and a `routing_preference` knob (`TRAFFIC_AWARE`, `TRAFFIC_AWARE_OPTIMAL`, `TRAFFIC_UNAWARE`) to trade accuracy versus cost.
* **Dual-format output** — every result returns both raw values (`duration_seconds`, `distance_meters`) and pre-formatted human strings (`duration_text` like `"1h 12m"`, `distance_text` like `"34.8 mi"`), with the unit system controlled by a `units` parameter (`imperial` by default, `metric` available).
* **Safe-by-default caps** — per-call limits exist as cost guardrails. The tool returns a clear error and recommends paired mode whenever a matrix call would be wasteful for the requested result shape.
**Why this matters:**
* **Real geo intelligence in agent loops** — every customer-facing agent can answer scheduling, dispatch, and routing questions ("which technician is closest to this work order?", "if I leave at 4pm Thursday, when do I arrive?") without anyone provisioning a Google Cloud account.
* **Bulk-first design** — fleet and territory workflows that need 50–100 lookups per run no longer need to fan out one agent call per pair. One tool call returns the full set, sorted however the agent asks.
* **No CASA, no consent screen, no friction** — this is the cheapest possible delivery model. The Maps Platform doesn't use OAuth; agents just call the tool.
Available now on every agent. Attach **Compute Route** to your agent in the tool picker and it works on first call.
### Google Analytics 4 for Your Agents
Aster now ships with a **Google Analytics 4 integration** — admins click **Connect** on the Google Analytics card under **Control Hub > Integrations**, sign in with a Google account that has access to their GA4 properties, and every agent in the org gains live access to traffic, engagement, conversion, and revenue data via the official GA4 Data API.
Like the Google Search Console integration, this uses standard user-OAuth (no service accounts, no JSON keys, no GCP project required of the customer). The connect flow shares a single Aster-side OAuth client with Google Search Console — adding more Google integrations later (Calendar, Sheets, Docs, etc.) will reuse the same plumbing.
**What's included:**
* **One-click OAuth setup** — Connect button, sign in with the Google account that owns/has access to the property, done. Refresh tokens are durable (the Aster OAuth app is in Google's production audience), so connections persist indefinitely as long as the user doesn't revoke access.
* **Two new agent tools:**
* **List GA4 Properties** — discover every GA4 account and property the connected Google account can read, via the Admin API's `accountSummaries` endpoint. Agents grab the right property ID at runtime instead of hard-coding it.
* **Run GA4 Report** — full-fidelity wrapper around the Data API's `runReport`. Supply a property ID plus the metrics you want and (optionally) dimensions, dimension/metric filters, ordering, pagination, and one or two date ranges (use two with `name` for period-over-period comparison). Returns flattened rows where each row is `{dim1: value, ..., metric1: number, ...}` — far easier for an LLM to reason about than the API's raw `dimensionValues`/`metricValues` arrays. Common dimensions and metrics are listed in the tool description so agents pick valid names on the first try.
* **Connection management modal** — see the connected Google account email, when the integration was authorized, and how long ago. Reconnect or disconnect from the same modal.
**Why this matters:**
* **Live engagement data in any agent** — ask "what were our top 10 landing pages last month by sessions?" or "compare conversion rate by source between this quarter and last" and get answers grounded in real GA4 data, not a stale CSV export.
* **Cross-property workflows** — agencies and multi-brand companies can query across every GA4 property the connected account can see in one tool call.
* **No per-customer GCP setup** — the customer's only setup is logging into Google. No project creation, no API enablement, no service-account email to add to property permissions.
Available now on all orgs. Connect under **Control Hub > Integrations > Google Analytics**.
### Bring Your Own Resend — Agents Send From Your Verified Domain
The **send\_email** tool now supports **Bring Your Own Resend**. Admins paste a Resend API key and a verified from-address under **Control Hub > Integrations > Resend**, and every agent in the org sends email from your domain — your deliverability, your reputation, your sender authentication. Without a connection, agents continue to send through the default Aster sender as before; nothing breaks.
As part of the same release, **send\_email** was ported from the Modal Python runtime to the Vercel JS tool runtime alongside the rest of the integration tools. Same tool name, same parameters, faster cold-start.
**What's included:**
* **One-screen setup** — paste a Resend API key (server-side, full-access with Send + Domains permission) and the from-address agents should send as. The address must live at a domain you've already verified in Resend (e.g. `ai@notifications.yourcompany.com`).
* **Test-send on first connect** — optionally provide a test recipient and the integration sends a one-time validation email to it before saving the credentials. Catches a bad key or unverified domain immediately, rather than at the first agent run.
* **Domain validation** — the connect endpoint cross-checks your from-address's domain against Resend's `/domains` API and rejects setup if it isn't verified.
* **Per-agent sender naming** — the display name on every outbound is the agent's own name (`"Acme Sales Agent "`), so recipients see who's writing them without manual config per agent.
* **Graceful fallback** — orgs without a Resend connection get the legacy Aster-sender behavior automatically. No agent config change needed.
**Why this matters:**
* **Deliverability under your control** — emails go through your verified domain with your SPF/DKIM/DMARC, not a shared Aster sender. Reply-tos route to you. Bounces and reputation are yours.
* **Compliance & trust** — many B2B prospects/customers see "from a domain we know" instead of "from an unfamiliar sender" — sometimes the difference between a reply and a junk folder.
* **Drop-in for existing agents** — every agent already using the `send_email` tool starts using your Resend account the moment you connect, without prompt changes.
Available now on all orgs. Connect under **Control Hub > Integrations > Resend**.
### Egnyte File Storage — Search, Read, and Manage Files in Agent Loops
Aster now ships with an **Egnyte integration** — connect an Egnyte tenant under **Control Hub > Integrations > Egnyte** and every agent in the org gets a Box-style toolkit for files: search, browse, read, upload, organize. Useful any time customer source-of-truth files live in Egnyte instead of (or alongside) Box, Drive, or SharePoint.
Auth is OAuth 2.0 against the customer's Egnyte tenant. Tokens are persisted in the org's connection row — refresh handling happens server-side, no in-process state to lose on a cold start.
**What's included:**
* **Admin-connected setup** — admins paste the Egnyte domain (e.g. `yourcompany.egnyte.com`), an API Key + Shared Secret from a registered Egnyte API application, and a service-account username/password. The integration handles token exchange and refresh from there.
* **Nine new agent tools** covering the full file-ops surface:
* **Egnyte Search** — filter by query string, path scope, file extension, modifier, and modified-since, returning files + folders that match.
* **Egnyte List Folder** — list children of a folder by path, with pagination and a flag for including subfolder counts.
* **Egnyte Read File** — pull a file's text/PDF/Office contents into the conversation for the agent to reason over.
* **Egnyte Get File Info** — metadata for a single file (size, modified time, permissions, version history) without downloading the body.
* **Egnyte Download File** — bring a file back as a downloadable attachment on the agent's response.
* **Egnyte Upload File** — write a new file to a folder path, from inline content or a generated artifact earlier in the conversation.
* **Egnyte Create Folder** — make a folder (with `mkdir -p` semantics).
* **Egnyte Move Item** — rename or relocate a file/folder.
* **Egnyte Copy Item** — duplicate a file/folder to a new path.
* **Token persistence in the DB** — OAuth tokens live on `organization_data.connections.egnyte`, not in an in-process map. Survives Vercel cold starts and concurrent invocations cleanly.
**Why this matters:**
* **Egnyte-first companies are first-class** — the Box integration was great if you used Box; the Drive scopes are coming. Egnyte fills the gap for the (large) cohort that runs on Egnyte as their content layer.
* **Reads + writes in one place** — agents can ingest a contract, redline it via execute\_python, and write the result back into the correct customer folder in one run.
* **Composable with the rest of the toolkit** — pair with `search_knowledge_base`, `execute_python`, and `write_to_knowledge_base` to build pipelines that pull from Egnyte, transform, and either return to Egnyte or land in a KB for later semantic search.
Available now on all orgs. Connect under **Control Hub > Integrations > Egnyte**.
### Gemini 3.5 Flash
**Gemini 3.5 Flash** (`google:gemini-3.5-flash`) is now available in Aster Agents. Google's latest lightweight model — fast, cheap, and surprisingly capable for structured extraction, summarization, and tool-calling workloads where latency matters more than raw reasoning depth.
Select **gemini-3.5-flash** from the model picker in **Control Hub > Edit Agent**.
### Latest Frontier Models, and a Spec for the Model String
The [Choosing a Model](/features/choosing-a-model) page now reflects the two latest flagships available on Aster — **Claude Opus 4.7** (`anthropic:claude-opus-4-7`) and **GPT-5.5** (`openai:gpt-5.5`) — and adds a dedicated **Model Identifiers** reference table.
**What's included:**
* **Refreshed provider table** — Anthropic row now lists Opus 4.7 alongside Opus 4.6, Sonnet 4.6, and Haiku 4.5. OpenAI row lists GPT-5.5 alongside the GPT-5.4 family.
* **Model Identifiers table** — the literal `provider:model-id` strings the platform expects (e.g. `anthropic:claude-sonnet-4-6`, `openai:gpt-5.5`, `azure:`), for every supported chat model. Bare display names like `"claude-sonnet-4-6"` are not accepted by the runtime — the full `provider:` prefix is required.
**Why this matters:**
* **Discoverability** — agents that consult the docs to pick a model (including those built via the Builder Agent) now see the latest options and pick a syntactically valid identifier on the first try, instead of guessing display names.
* **API/tool correctness** — code calling the [Create or Update an Agent](/api-reference/agents/create-or-update-an-agent) endpoint or invoking the `manage_agents` tool now has an unambiguous reference for the `model` field.
No platform changes — these models were already available in **Control Hub > Providers**; the docs simply caught up.
### Aspire Integration — Read-Only Access to Your Field Service Data
Aster now ships with an **[Aspire](https://youraspire.com) integration** for landscape and field service companies. Connect your Aspire tenant under **Control Hub > Integrations > Aspire** and your agents can query Properties, Opportunities, Work Tickets, Invoices, Routes, and more — all read-only.
**What's included:**
* **Client Credentials setup** — admins paste an Aspire Client ID and Secret (generated under Account Settings > API Keys in Aspire) and the integration handles OAuth token exchange and refresh from there.
* **Read-only by design** — agents can query data across Aspire's REST API but cannot create, update, or delete records. This is intentional for the initial release.
Available now on all orgs. Connect under **Control Hub > Integrations > Aspire**.
### Google Search Console for Your Agents
Aster now ships with a **Google Search Console integration** — admins click **Connect** on the Google Search Console card under **Control Hub > Integrations**, sign in with a Google account that has access to their Search Console properties, and every agent in the org can query clicks, impressions, CTR, and average position across those properties.
The connect flow uses standard user-OAuth — no service accounts, no JSON keys, no GCP project setup required of the customer. Aster's OAuth app is in Google's production audience, so refresh tokens are durable; connections persist indefinitely unless the user revokes access.
**What's included:**
* **One-click OAuth setup** — Connect button, sign in with a Google account, done. Tokens never leave the server and are redacted in API responses; access tokens auto-refresh behind the scenes.
* **Two new agent tools:**
* **List Search Console Sites** — discover every property the connected Google account can read, plus the permission level on each (Full, Restricted, Owner). Agents grab the exact `siteUrl` at runtime instead of hard-coding it.
* **Search Analytics Query** — the workhorse. Pull clicks, impressions, CTR, and average position for a property over any date range (data goes back \~16 months). Optionally group by `query`, `page`, `country`, `device`, `date`, or `searchAppearance` — single dimension or any combo. Filter groups support `contains` / `equals` / `notContains` / regex on any dimension, with AND-of-groups / AND-within-group logic for arbitrary precision. Up to 25,000 rows per call with offset-based pagination, plus a `dataState` switch to choose between live (fresh, may revise) and finalized data.
* **Connection management modal** — see the connected Google account email, when the integration was authorized, and how long ago. Reconnect or disconnect from the same modal.
**Why this matters:**
* **Live SEO data in any agent** — ask "what were our top 20 queries last month, ranked by clicks?", "which pages lost impressions week-over-week?", or "give me CTR by device for our pricing page over the last 90 days" and get answers grounded in real Search Console data, not a stale spreadsheet.
* **Cross-property workflows** — agencies and multi-site companies can pull data across every property the connected account can see in a single tool call, then have the agent aggregate or compare.
* **Foundation for more Google services** — this integration shares its OAuth scaffolding with the Google Analytics 4 integration that landed the following week, and is the same plumbing any future Google service (Calendar, Sheets, Docs, etc.) will plug into. One Aster OAuth client, one redirect URI, one verification submission with Google.
Available now on all orgs. Connect under **Control Hub > Integrations > Google Search Console**.
### Aster Agents Now Speaks Asana
Aster ships with a new [Asana integration](/integrations/asana) — connect with a Personal Access Token and your agents can search, create, update, and manage tasks, projects, sections, and comments across your entire Asana workspace.
**What's included:**
* **PAT-based setup** — admins paste an Asana Personal Access Token under **Control Hub > Integrations > Asana** and every agent in the org can use it. Service Account tokens (Asana Enterprise) are also supported for org-wide access that doesn't depend on a single user.
* **Three new agent tools:**
* **Get Asana Schema** — discover workspaces, teams, projects, users, tags, sections in a project, and the custom fields applied to a project. Agents look up gids at runtime instead of hard-coding them.
* **Search Asana** — filter tasks across a workspace by free-text plus a deep filter set: assignee/projects/sections/tags/followers any-or-not, completion, blocked/blocking, due/start/completed/created/modified date ranges (with `is_set` variants for "tasks with no due date"), engagement (commented-on-by, liked-by), subtask flag, attachment flag, and any custom field value (8 operators including `is_set`, `between`, `any`, `not`). A `rawParams` escape hatch covers anything else Asana accepts.
* **Manage Asana Records** — single tool with an `action` enum that handles both CRUD on tasks/projects/sections/comments **and** all task-relationship operations: add/remove project (with section placement), move between sections, add/remove tags, add/remove followers, add/remove dependencies and dependents, and `set_parent` for subtask linkage. One tool slot, no juggling.
* **Custom-field-aware writes** — set task custom fields on create/update via `data.custom_fields` keyed by gid, including enum option gids and multi-enum arrays.
**Why this matters:**
* **Live work data in any agent** — ask "what's overdue and unassigned in the Q2 Roadmap project?" or "summarize every task closed last week" and get answers grounded in real Asana state, not stale exports.
* **Two-way automation** — agents that read can also write: turn an inbound email into a task in the right section with the right custom fields, post comments on stale tasks during a triage routine, mark a batch of tasks complete from a Slack approval, etc.
* **Org-level connection** — every agent in your workspace can use Asana once an admin connects it once. No per-agent setup.
* **One tool, fewer tokens** — keeping CRUD and relationship operations on a single tool means agents pick a tool once and dispatch on `action`, instead of choosing between two similar-sounding manage tools.
**Pagination caveat:** Asana's tasks-search endpoint returns up to 100 results per call **with no cursor**. To fetch more, agents narrow filters — the tool returns `hasMore: false` to make this honest.
Available now on all orgs. See the [Asana setup guide](/integrations/asana) to get started.
### One-click QuickBooks Online for Your Agents
Aster now ships with a [QuickBooks Online integration](/integrations/quickbooks) — connect your QBO company file in one click and your agents can query accounting data and run financial reports.
**What's included:**
* **One-click OAuth setup** — admins click Connect on the QuickBooks card in **Control Hub > Integrations**, authorize via Intuit, and they're done. No API keys, no refresh tokens to manage. Access tokens auto-refresh every 60 minutes behind the scenes.
* **Four new agent tools:**
* **QuickBooks Query** — run QBO `SELECT` statements against any entity (Customer, Vendor, Invoice, Bill, Payment, Account, Item, SalesReceipt, CreditMemo, Deposit, Transfer, JournalEntry, etc.) with automatic pagination.
* **QuickBooks Get Schema** — agents discover available entities and field names at runtime, with example queries for each.
* **QuickBooks Get Report** — pull P\&L, Balance Sheet, Cash Flow, AR/AP aging, customer/vendor balances, item sales, and \~25 other built-in QuickBooks reports. Supports relative date macros like `This Fiscal Year-to-date` and `Last Quarter`.
* **QuickBooks Manage Records** — create, update, delete, or void records across 17 writable QBO entities. Single tool with an `action` enum, sparse update semantics (only fields you pass change; everything else is preserved), and automatic SyncToken handling.
* **Connection management modal** — see the connected Company ID, when the integration was authorized, and how long until the OAuth grant expires (with a warning when it's within 14 days). Reconnect or disconnect from the same modal.
**Why this matters:**
* **Live financial data in any agent** — ask "what was our P\&L last quarter?" or "which customers have outstanding balances over \$5k?" and get answers grounded in real QuickBooks data, not stale exports.
* **Two-way automation** — agents that read can also write: turn an inbound email into an Invoice, post month-end JournalEntries from a Slack approval, void a duplicate Bill detected during AP review, etc.
* **Org-level connection** — every agent in your workspace can use QuickBooks once an admin connects it once. No per-agent setup. To make an agent read-only, simply don't enable Manage Records on it.
* **OAuth-grade security** — tokens never leave the server, are redacted in API responses, and the consent flow is CSRF-protected with a dedicated signing secret.
Available now on all orgs. See the [QuickBooks Online setup guide](/integrations/quickbooks) to get started.
### Sub-Agent Threads in the Sidebar
Multi-agent runs (anything that uses [Call Agent](/tools/call_agent), e.g. a Master Portfolio Agent that fans out to dozens of company-specific sub-agents) now expose every sub-agent's chat directly from the **Conversation History** in the left sidebar. Previously the sub-agent threads existed in the database but were not surfaced — a single multi-agent burst could push every other recent conversation off the sidebar.
**What's included:**
* **`+N` badge** on each root conversation in the sidebar showing the count of sub-agent threads it spawned (recursive — covers the full descendant tree, not just direct children).
* **"Sub-threads (N)" entry** in the row's actions menu (the ⋯ icon that appears on hover) opens a modal listing every sub-agent.
* **Scannable layout** — the modal renders agent names in a 1, 2, or 3-column grid based on count, so a 30-sub burst fits in one screen without internal scroll. A filter input appears above 8 items to jump to a specific sub-agent by name.
* **One-click drill-in** from the modal into any sub-agent's full conversation, including its tools, attachments, and reasoning trace.
**Why this matters:**
* **Auditability** — when an agent delegates, you can now see exactly which sub-agents ran and what each one did, without needing to know the parent conversation's URL.
* **Sidebar stays useful** — heavy multi-agent users no longer get their recent-chats list dominated by a single fan-out.
Available now on all orgs. No configuration needed.
### Better Image Generation — GPT-Image-2
The built-in image generation tool now uses **GPT-Image-2**, OpenAI's latest image model. It ships O-series reasoning before rendering, supports native up-to-4K resolution, and handles multi-turn editing — modify a previously generated image by describing what to change, rather than starting from scratch.
All agents with image generation enabled get the upgrade automatically. No configuration changes needed.
### Preview Email Body Before Sending
The `send_email` tool's output now renders the HTML body in a **sandboxed iframe preview** directly in the chat UI. Agents and users can see exactly what the recipient will receive — formatted text, links, layout — before the email goes out, instead of reading raw HTML.
### Scheduled Tasks: One Tool, Five Actions
The four separate scheduling tools (`schedule_task`, `list_scheduled_tasks`, `edit_scheduled_task`, `delete_scheduled_task`) have been **consolidated into a single `manage_scheduled_tasks` tool** with an `action` enum — matching the pattern used by `manage_agents` and `manage_knowledge_bases`.
**What's included:**
* **`action: "create"`** — schedule a new recurring task (cron expression + prompt)
* **`action: "list"`** — list all tasks in the org, with optional `nameFilter` and `enabledFilter`
* **`action: "get"`** — fetch a single task by `taskId`, **including recent execution history** with status badges and links to the conversation that ran for each execution
* **`action: "update"`** — edit name, description, prompt, schedule, or enabled state
* **`action: "delete"`** — remove a task
**Why this matters:**
* **Visibility into past runs** — clicking an execution row jumps straight into the conversation thread that ran, so you can audit what the agent actually did each time the task fired
* **Cleaner tool surface** — one tool slot per agent instead of four
* **Consistent with other manage tools** — same `action` pattern as `manage_agents` and `manage_knowledge_bases`
All existing agents using the previous tools have been migrated automatically. No configuration changes needed.
### ElevenLabs Audio Tools
Two new tools bring ElevenLabs' voice synthesis and transcription to your agents. Generate spoken audio from text and transcribe audio or video from anywhere on the web or from files already in the conversation.
**What's included:**
* **[elevenlabs\_text\_to\_speech](/tools/elevenlabs_text_to_speech)** — convert text to speech with a curated enum of 10 voices (`rachel`, `sarah`, `jessica`, `charlotte`, `lily`, `george`, `brian`, `daniel`, `will`, `charlie`), plus a `voice_id` escape hatch for custom cloned voices. Defaults to the latest `eleven_v3` model; configurable `output_format` (MP3, PCM, µ-law). The generated audio file is attached to the conversation for playback, download, or further processing.
* **[elevenlabs\_speech\_to\_text](/tools/elevenlabs_speech_to_text)** — transcribe audio with ElevenLabs Scribe (`scribe_v2` by default). One `audio_source` parameter accepts either an `r2://` path (a file already attached to the conversation — e.g. from TTS output or a user upload) or any HTTPS URL (cloud storage, YouTube, TikTok, podcasts, meeting recordings — up to 2GB). Returns the transcript with word-level timestamps, speaker diarization, audio-event tagging, and auto language detection.
**Common use cases:**
* Generate voiceovers for videos, podcasts, or onboarding flows directly inside an agent
* Round-trip text → audio → transcript for QA or accessibility alternates
* Transcribe recordings that already live in a knowledge base or prior conversation attachment — no re-upload needed
* Pull searchable transcripts from publicly hosted audio/video content
Enable them on any agent from the **Audio** section of the tool picker in **Control Hub > Edit Agent**.
### Airtable Integration
New integration connecting agents to your Airtable bases with three tools covering dynamic schema discovery, flexible record search, and full CRUD.
**What's included:**
* **airtable\_get\_schema** — list accessible bases, then inspect tables, fields (with select choices), and views inside any base
* **airtable\_search** — query records three ways: free-text `searchTerm` across named fields, raw `filterByFormula` for full formula-language filtering, or plain list with sort/view/fields. Supports offset pagination.
* **airtable\_manage\_records** — create, read, update, upsert, and delete records. Batch writes capped at 10 records per call (Airtable's API limit). `typecast` flag coerces human-readable strings into existing select options, numbers, and dates.
**Authentication:** Uses Airtable [Personal Access Tokens](https://airtable.com/create/tokens) (starts with `pat`). Tokens are scoped per-base with `data.records:read`, `data.records:write`, and `schema.bases:read` scopes. Grant the PAT access to exactly the bases the agent should reach.
See the [Airtable integration docs](/integrations/airtable) for setup details.
### GPT-5.5
**GPT-5.5** is now available in Aster Agents on the day of [OpenAI's release](https://openai.com/index/introducing-gpt-5-5/). OpenAI's smartest model yet, with major gains on coding, multi-step workflows, and autonomous research.
**Specs:**
* 1M context window
* 128K max output tokens
* Vision support
* Reasoning token support
* $5 / $30 per 1M input / output tokens
Select **gpt-5.5** from the model picker in **Control Hub > Edit Agent**. New OpenAI agents onboarded from today default to GPT-5.5.
### Sub-Agents Now Inherit Parent Conversation Attachments
When an agent delegates work via [Call Agent](/tools/call_agent), the sub-agent now **automatically receives the parent conversation's file attachments** — both user uploads and files generated earlier in the conversation — as part of its first message.
**Why this matters:**
* Build multi-agent pipelines where a file flows from one specialist to another (e.g., coordinator → scope writer → estimator)
* No need to re-upload or re-reference files when switching between agents mid-workflow
* Generated files (Python sandbox outputs, tool-produced docs) are passed through alongside user uploads
Works automatically on any agent that uses the `call_agent` tool — no configuration changes needed.
### Token Cost Estimates on the Dashboard
The **Tokens Used** card on the Dashboard now includes an **estimated dollar cost** in its hover tooltip, calculated from a per-model price breakdown across both agent conversations and knowledge base document processing.
* Per-model input and output tokens are multiplied by the model's listed price
* If any tokens came from a model without listed pricing, the estimate is flagged with "(some models unpriced)"
* Applies to both the stat card and the token trend chart
### Google Sheets Tools Temporarily Removed
The `read_google_sheet` and `append_google_sheet` tools have been **removed**. The per-user OAuth flow we'd been using couldn't support unattended runs — so scheduled tasks and agent-initiated calls were silently failing.
We'll bring Sheets back with a long-lived token flow so it works across all execution contexts. In the meantime, the [Call API](/tools/call_api) tool can hit the Google Sheets REST API directly as a workaround.
### Search Places and Read Reviews from Google Maps
Two new research tools backed by the SearchAPI Google Maps engines:
* **[search\_google\_maps](/tools/search_google_maps)** — find places by name, category, or query. Returns `place_id`, address, rating, phone, website, hours, and more
* **[get\_business\_reviews](/tools/get_business_reviews)** — pull reviews for a place. Accepts either a `place_id` directly or a business name (with optional location) — the tool looks up the place automatically
Useful for local lead research, competitive analysis, and review monitoring. Enable them on any agent from the tool picker in Control Hub.
### Cleaner Status for Files That Can't Be Indexed
Files that can't be indexed for semantic search (images, `.xlsb` workbooks, ZIP archives, and other unsupported types) no longer show a red **Failed** badge. They now get a blue **Stored** badge indicating the file is saved and accessible — it just isn't part of the searchable index.
Agents can still see and retrieve these files via knowledge base tooling; they're simply skipped when running a semantic search.
### Invite External Users Safely with the Guest Role
You can now invite clients, partners, and other external users into your organization and let them chat with your agents **without** exposing the Control Hub, knowledge bases, skills, or anything else you've set up behind the scenes.
**Why this matters:**
* Share a private agent with a client without them seeing how it was built
* Onboard contractors for a short engagement and remove their access cleanly when it ends
* Let external stakeholders interact with workflows your team curates, without risk of them touching the data or configuration
**What guests can do:**
* ✅ Chat with any agent in your organization
* ✅ See their own conversations and dashboard
* ❌ Open the Control Hub or any of its pages — they're redirected back to chat
* ❌ List, view, or modify knowledge bases, skills, scheduled tasks, or agent configs
* ❌ Discover which agents or data sources exist outside of what they interact with
**How to invite a guest:**
1. In Clerk, invite the user to your organization and select the **Guest** role.
2. They sign in and land directly on the chat — Control Hub is hidden from their navigation.
3. Remove them from the organization when the engagement ends — their access is revoked immediately.
See the updated [Roles & Permissions guide](/settings/roles-permissions) for the full role matrix and common patterns (agencies, embedded external users, development teams).
### Search Knowledge Base Gets Allow-All Toggle
Mirroring the [Call Agent allow-all mode](/tools/call_agent), the Search Knowledge Base tool now has an **"Allow all knowledge bases in organization"** toggle. When enabled, the agent can search any KB in the org without a curated allowlist.
* Pairs naturally with the new Manage Knowledge Bases tool for agents that discover KBs at runtime
* The strict allowlist still applies when the toggle is off
* [Write to Knowledge Base](/tools/write_to_knowledge_base) is unchanged — writes still require an explicit allowlist
Toggle it in **Control Hub > Edit Agent > Tools > Search Knowledge Base**.
### Manage Knowledge Bases Tool
New platform tool that lets agents **discover knowledge bases at runtime**, mirroring the [Manage Agents](/tools/manage_agents) pattern.
**What's included:**
* **List** — agents can see every KB in the org (id, name, description, file counts, extraction schema, embedding model)
* Pass the discovered `id` directly to `search_knowledge_base` — no curated allowlist required when paired with the KB allow-all mode
**Why it matters:** Orchestrator agents can decide *which* knowledge base to search based on the user's question, rather than always querying a fixed set.
Enable `manage_knowledge_bases` from the agent's tool list in **Control Hub > Edit Agent > Tools**.
### Claude Opus 4.7
**Claude Opus 4.7** is now available in Aster Agents on the day of Anthropic's [general availability release](https://www.anthropic.com/claude/opus). It's Anthropic's most powerful GA model, with stronger performance across coding, vision, and complex multi-step tasks.
**What's new:**
* **Better software engineering and instruction-following** — higher-quality output on real-world work
* **Improved vision** — sees images at greater resolution
* **1M context window** at standard pricing (same as Opus 4.6)
* **Same price as Opus 4.6** — no premium for the upgrade
Select **claude-opus-4-7** from the model picker in **Control Hub > Edit Agent** to use it for any agent.
### HubSpot CRM Integration
New integration connecting agents to your HubSpot CRM with four tools covering search, CRUD, associations, and schema discovery.
**What's included:**
* **hubspot\_search** — filter and sort any object type with AND/OR logic, property selection, and pagination
* **hubspot\_manage\_records** — create, read, update, upsert, and delete records on contacts, companies, deals, tickets, and more
* **hubspot\_manage\_associations** — list, create, and remove links between records
* **hubspot\_get\_schema** — discover properties, types, and picklist values for any object
**Authentication:** Uses HubSpot [Service Keys](https://developers.hubspot.com/docs/apps/developer-platform/build-apps/authentication/account-service-keys) (recommended) or Private App tokens. Service Keys are scoped, rotatable, and not tied to an individual user.
See the [HubSpot integration docs](/integrations/hubspot) for setup details.
### Call Agent: Allow All Mode
The Call Agent tool now supports an **"Allow all agents in organization"** mode. When enabled, the agent can call any agent in the org without needing a curated allowlist — ideal for orchestrator agents that use [Manage Agents](/tools/manage_agents) to dynamically discover and delegate to other agents at runtime.
**What changed:**
* New toggle in Control Hub: **Allow all agents in organization** (on by default for new configurations)
* When toggled off, the existing agent selector modal appears for picking specific callable agents
* Agents with existing allowlists are unaffected — the allowlist still applies when configured
See the updated [Call Agent tool docs](/tools/call_agent) for details.
### Manage Agents Tool
New platform tool that gives agents **self-awareness** of the organization's agent ecosystem. Agents can now list, create, and update other agents — all from within a conversation.
**What's included:**
* **List** — discover all agents in the org with names, IDs, descriptions, tools, tags, and stage. The calling agent is marked with `isCurrentAgent: true`
* **Create** — build new agents with a name, system prompt, model, and tools. Pass tool names as simple strings (e.g., `["search_google", "scrape_url"]`) and the platform resolves them to full schemas
* **Update** — modify an existing agent's description, system prompt, model, tools, or stage. Existing tool configs (like callable agent IDs) are preserved
**Why it matters:** Agents can now discover available agents before using `call_agent`, rather than requiring hardcoded agent names. Combined with create, this enables agents to build and orchestrate other agents dynamically.
See the [Manage Agents tool docs](/tools/manage_agents) for full details.
### Salesforce CRM Integration
New integration connecting agents to your Salesforce CRM with three tools:
* **salesforce\_query** — execute SOQL queries with auto-pagination (up to 12k records across 5 pages)
* **salesforce\_manage\_records** — create, read, update, upsert, and delete records on any standard or custom object
* **salesforce\_get\_schema** — discover available objects and inspect field metadata to inform queries
Authentication uses username + password + security token via SOAP login — no Connected App required. Sessions are cached with automatic refresh on expiry.
See the [Salesforce integration docs](/integrations/salesforce) for setup details.
### Parallel Enrich Tool
New async tool for **batch enrichment** of entities (contacts, companies, leads) via AI-powered web research. Submit up to 1,000 entities in a single call — results are delivered asynchronously via webhook as each entity completes.
**Key features:**
* Two research depth tiers: **base** (quick lookups) and **core** (thorough research including email, phone, and social profiles)
* Custom `task_instructions` for tailored research guidance per batch
* Results stream back into the conversation as they arrive — no polling needed
See **Parallel Enrich** for full details.
### Public Skills
Skills can now be **public** — visible and usable by all organizations on the platform. Public skills appear alongside org-owned skills in the Skills list and agent configuration, marked with a "Public" badge.
**How it works:**
* Public skills are platform-managed and read-only for all organizations
* They appear automatically in every org's skill list and agent skill picker
* Agents treat public and org-owned skills identically at runtime — no behavioral difference
* If an org creates a skill with the same name as a public skill, the org's version takes precedence
**What's read-only:**
* Public skills cannot be edited, deleted, or have files uploaded by org users
* Download is still available for reference
### SQL Server Integration
Agents can now query **Microsoft SQL Server** databases directly with the new `mssql_sql` tool. Execute T-SQL queries with automatic CSV export for large result sets, token-aware inline truncation, and secure connection configuration via organization settings.
See the [SQL Server integration docs](/integrations/sql-server) for setup details.
### New OpenAI Models & Deprecations
Added **GPT-5.4 mini** and **GPT-5.4 nano** — OpenAI's latest efficient reasoning models.
**New models:**
* **GPT-5.4 mini** — strong reasoning in a fast, cost-effective package (400k context, 128k output)
* **GPT-5.4 nano** — cheapest GPT-5.4-class model for high-volume tasks like classification and data extraction (400k context, 128k output)
**Deprecated & migrated:**
* `o3` agents migrated to **GPT-5.4**
* `o4-mini` agents migrated to **GPT-5.4 mini**
* `gpt-5-mini` and `gpt-5-nano` removed in favor of their 5.4-series replacements
### Agents Can Browse the Web
New tool: **Browser Use** — gives any agent a real browser sandbox powered by Firecrawl. Agents navigate to URLs, read page content via an accessibility tree, click elements, fill forms, scrape content as markdown, and download files — all within a sandboxed session that persists across tool calls in a conversation.
**What's included:**
* **Core commands** — `open` (navigate), `snapshot` (get page accessibility tree with interactive element refs), `click`, `fill`, `scrape` (extract full page as markdown)
* **Session persistence** — a `session_id` is returned on first use and reused across subsequent calls, so the agent maintains browser state (login sessions, multi-page flows, pagination)
* **File downloads** — files downloaded in the browser are automatically uploaded to R2 and attached to the conversation as downloadable artifacts
* **Isolation** — each conversation gets its own browser profile; sessions are ephemeral and not shared across conversations
**Common use cases:**
* Scrape structured data from websites that don't have APIs
* Fill out web forms on behalf of users
* Navigate authenticated portals after logging in
* Download reports or files from web dashboards
Enable `browser_use` on any agent from the tool picker in **Control Hub > Edit Agent**.
### Meta Ad Library — Political Ads, Spend Data, and Demographics
New integration: **Meta Ads** — connect a Meta access token and your agents can search the Meta Ad Library for political and social-issue ads across any country. Returns ad creative, spend ranges, impressions, demographic breakdowns, and funder information.
**What's included:**
* **search\_meta\_ads** — filter by country, search terms, advertiser, platform (Facebook/Instagram/Messenger), active status, media type, and date range. Results sorted by spend descending with an optional `min_spend` filter to cut noise. Supports compact and detailed response modes.
* **System User tokens recommended** — system user access tokens don't expire (unlike 60-day user tokens), making this reliable for scheduled monitoring agents.
Connect under **Control Hub > Integrations > Meta Ads**. Requires a Meta access token with `ads_read` permission.
### Enhanced Conversation Feedback
The feedback experience at the end of conversations has been upgraded. After clicking thumbs up or thumbs down, a dialog now appears where users can share more detail about their experience.
**What's new:**
* **Category tags** (for negative feedback): quickly tag issues as Inaccurate, Not relevant, Too slow, Confusing, Incomplete, or Other
* **Comment field**: optional free-text to explain what went well or what went wrong
* **Skip option**: users can still submit a quick thumbs up/down without adding details
Feedback categories and comments are visible to admins in the Dashboard conversation view. All feedback data is stored per user per conversation — submitting again updates the existing entry rather than creating duplicates.
### Upload and Download Skills as Zip Packages
Skills now support **zip upload and download** — upload a zip containing `SKILL.md` and any bundled files (scripts, templates, references) to create or update a skill in one step, and download any skill as a zip for backup, sharing, or version control.
* **Zip upload** — drag a `.zip` file onto the skill editor or use the upload button. Nested folder structures are preserved. `SKILL.md` is extracted as the skill instructions; everything else becomes bundled files.
* **Zip download** — click the download button on any skill to get a complete package of its instructions and files.
### 1M Context Window — Now Standard
**Claude Opus 4.6** and **Claude Sonnet 4.6** now include the full **1M context window at standard pricing** — no separate model variant required. This follows Anthropic's [general availability announcement](https://www.anthropic.com/news/1m-context-ga).
**What changed:**
* The `(1M)` model variants have been merged into the standard models
* No more premium pricing for long-context requests — same per-token rate at any context length
* Existing agents using the 1M variants have been automatically migrated
If your agent was previously using `claude-opus-4-6 (1M)` or `claude-sonnet-4-6 (1M)`, it now runs on the standard model with the same 1M context window.
### Connect Any Remote MCP Server
MCP (Model Context Protocol) integration has been upgraded to support **remote hosted servers** with both **Streamable HTTP** (recommended) and **SSE** transports.
**What's new:**
* Add any remote MCP server by URL — no pre-installed packages required
* Streamable HTTP transport support (the latest MCP standard)
* Simple configuration: name, URL, and optional auth headers
* Auth headers are masked in the UI and redacted in API responses
**How to use:**
1. Go to **Control Hub > Integrations** and scroll to **MCP Servers**
2. Click **Add Server**, enter the server URL and any required auth headers
3. Assign the server to agents in **Control Hub > Edit Agent > MCP Servers**
Agents will automatically discover and use tools provided by connected MCP servers during conversations.
### Campaign Finance Data for Your Agents
New integration: **OpenSecrets** — connect a FollowTheMoney API key and your agents can search campaign finance data, donor records, and lobbying activity across all 50 US states.
**What's included — five tools:**
* **opensecrets\_search\_contributions** — search individual and organizational contributions by state, year, office type, donor name, employer, and amount range
* **opensecrets\_search\_candidates** — find candidates by state, office, party, election year, and name
* **opensecrets\_candidate\_industries** — see which industries funded a candidate
* **opensecrets\_top\_donors** — top donors to a candidate or committee
* **opensecrets\_get\_entity** — look up details for any entity (candidate, donor, organization) by ID
Connect under **Control Hub > Integrations > OpenSecrets**. Requires a FollowTheMoney API key from [myFollowTheMoney.org](https://myFollowTheMoney.org).
### Add to Contacts Button
Agent and knowledge base email addresses now include an **Add to Contacts** button. Clicking it downloads a `.vcf` (vCard) file that works on any operating system — macOS, Windows, iOS, Android, and Linux — so you can save agent and KB email addresses to your contacts app in one click.
Available on agent email settings and knowledge base detail pages.
### On-Demand Agent Skills
Agents can now load specialized instructions and bundled files on demand using the new [Agent Skills](/features/skills) feature.
**How it works:**
* Create a skill with a `SKILL.md` file containing instructions and YAML frontmatter
* Upload bundled files (scripts, templates, references) alongside the instructions
* Assign skills to agents — they see only names and descriptions in their system prompt
* When a task matches, the agent calls `load_skill` to get the full instructions and file manifest
* Bundled scripts are pre-loaded into the Python sandbox at `/home/user/skills/{skill-name}/`
**Key features:**
* Progressive disclosure keeps system prompts lean
* Up to 20 bundled files per skill (scripts, templates, references)
* Name-based skill loading for natural agent interaction
* Full [API support](/api-reference) for programmatic skill management
Create and manage skills in **Control Hub > Skills**.
### Agents Can Save Files to Knowledge Bases
Agents can now **write directly to knowledge bases** during conversations using the new [Write to Knowledge Base](/tools/write_to_knowledge_base) tool.
**Two modes:**
* **Save attachments** — user-uploaded or tool-generated files from the conversation can be saved to a KB
* **Create new files** — agents can write inline content (summaries, reports, notes) as new KB files
**Built-in guardrails:**
* Configure which knowledge bases each agent can write to from the agent's tool settings
* 10 MB content size limit for inline writes
* Files are automatically processed, chunked, and indexed for future search
Enable the tool and select writable knowledge bases in **Control Hub > Edit Agent > Tools**.
### Send Notes and Updates to Knowledge Bases via Email
* **No Attachment Required**: Emails sent to a KB address without attachments now save the email body as a `.md` file
* **Rich Metadata**: Each file includes the subject, sender, and timestamp — searchable by your agents
* **Notes Storage**: Keep knowledge bases up to date by emailing interaction logs, meeting notes, or status updates
* **Same Pipeline**: Email body files go through the same chunking, embedding, and trigger pipeline as any other document
Works with the existing KB email addresses (`kb-{id}@updates.asteragents.com`). Emails with attachments continue to work as before.
### X Search and Web Search via Grok
Agents using xAI models now have access to two new **provider tools** powered by the Grok Responses API:
* **X Search** — search X (Twitter) for posts, users, and threads directly from your agent
* **Web Search** — search the web for current information in real time
Enable these tools on any agent using a Grok model from the agent's tool configuration.
### Legislative Tracking Tools
New [LegiScan integration](/integrations/legiscan) lets your agents search, analyze, and monitor legislation across all 50 US states and Congress.
**7 new tools:**
* **Search Bills** — find legislation by keyword, state, year, and status
* **Get Bill Details** — sponsors, history, votes, texts, and amendments
* **Get Bill Text** — read the actual text of a bill
* **Get Roll Call** — detailed vote tallies with individual legislator votes
* **Get Legislator** — look up legislator profiles
* **Monitor List & Set Monitor** — watchlist management for proactive tracking
Pair monitoring tools with [Scheduled Tasks](/features/scheduled-tasks) to automatically track bills and get notified of changes.
Connect your LegiScan API key in **Control Hub > Integrations** to get started.
### Box Integration — Search, Read, and Manage Files in Agent Loops
Aster now ships with a **Box integration** — connect your Box enterprise account under **Control Hub > Integrations > Box** and every agent in the org gets a full file toolkit.
**What's included — nine tools:**
* **box\_search** — search files and folders by name, content, or metadata across your Box enterprise
* **box\_list\_folder** — browse folder contents with pagination
* **box\_read\_file** — pull a file's text/PDF/Office content into the conversation for the agent to reason over
* **box\_get\_file\_info** — metadata for a file (size, modified time, shared links, version) without downloading the body
* **box\_download\_file** — bring a file back as a downloadable attachment on the agent's response
* **box\_upload\_file** — write a new file to a folder, from inline content or a generated artifact earlier in the conversation
* **box\_create\_folder** — create a folder at a given path
* **box\_move\_item** — rename or relocate a file/folder
* **box\_copy\_item** — duplicate a file/folder to a new location
**Authentication:** Uses Box's Client Credentials Grant (CCG) — admins paste a Client ID, Client Secret, and Enterprise ID from the Box Developer Console. Token refresh is handled server-side.
Available now on all orgs. Connect under **Control Hub > Integrations > Box**.
### Agents Can Now Run Much Longer
Agents are no longer limited by a single serverless execution window. When an agent needs more time to complete a complex task, Aster automatically **continues the work across multiple invocations** — seamlessly picking up where it left off.
**What this means:**
* Agents can now take **up to 500 tool-call steps** to complete a task (previously capped at 50)
* Long-running workflows that previously timed out or stopped silently will now run to completion
* The chat UI shows a "still working" indicator while your agent continues in the background
* If an agent does exhaust its limits, you'll see a clear message instead of a silent stop
No configuration changes needed — all agents benefit automatically.
### OpenAI's GPT-5.4 Models Now Available
Added support for **GPT-5.4**, **GPT-5.4 Pro**, and **GPT-5 Nano** — OpenAI's latest model family.
**New models:**
* `gpt-5.4` — 1M context window, 128K max output, vision support. Best all-around model for coding, reasoning, and agentic tasks.
* `gpt-5.4-pro` — Same specs as GPT-5.4 with deeper reasoning for tough problems.
* `gpt-5-nano` — 128K context, cost-optimized for high-throughput tasks like classification and instruction-following.
Select any of these from the model dropdown in your agent configuration.
### Test in Chat Button
Added a **Test in Chat** button to the agent configuration page. Quickly open a new chat with any agent — even ones hidden from the chat dropdown — directly from the Edit Agent screen.
### Human-in-the-Loop Questions
Agents can now **pause mid-conversation** to ask users interactive questions before continuing. The new [Ask User Question](/tools/ask_user_question) tool renders a rich UI with clickable option buttons, free text input, and multi-question pagination.
**Highlights:**
* Single or multiple questions in one tool call
* Predefined options with optional descriptions
* Client-side pagination for multi-question flows — all answers submitted together
* Free text input with or without predefined options
* Agent stops and waits for user response, then continues with full context
Add `ask_user_question` to any agent's tool list, then sync tools via **Control Hub > Sync Tools** or the API.
### Anthropic 1M Context Window Now Available
Added **1M context window** variants for Claude Opus 4.6, Sonnet 4.6, and Sonnet 4.5. These appear as separate models in the dropdown and use Anthropic's extended context beta to handle significantly larger documents and conversations.
**Available models:**
* `claude-opus-4-6 (1M)` — 1M input tokens, 128K max output
* `claude-sonnet-4-6 (1M)` — 1M input tokens, 64K max output
* `claude-sonnet-4-5 (1M)` — 1M input tokens, 100K max output
**Note:** Requests exceeding 200K tokens are charged at premium rates (2x input, 1.5x output). Requires Anthropic usage tier 4.
Select any `(1M)` model from the model dropdown in your agent configuration to start using it.
### Anthropic's Latest Sonnet Model Now Available
Added support for **Claude Sonnet 4.6**, the newest model in Anthropic's Claude 4.6 family.
**Specs:**
* 200K context window
* 64K max output tokens
* Extended thinking support
Select `claude-sonnet-4-6` from the model dropdown in your agent configuration to start using it.
### Redesigned Dashboard Metrics
The dashboard's six stat cards have been redesigned to surface more meaningful metrics, including knowledge base activity.
**New stat cards:**
* **Conversations** — now includes message count as subtext
* **Active Agents** — number of distinct agents used in the period
* **Active Users** — number of distinct users who had conversations
* **Tokens Used** — now includes document processing tokens alongside agent conversation tokens, with an input/output breakdown available on hover
* **Documents Processed** — total knowledge base files processed, with the number of knowledge bases shown as subtext
* **Feedback Score** — now includes total response count as subtext
**Analytics improvements:**
* Token Usage chart now shows **Agent Conversations** and **Document Processing** as two separate trend lines for clearer visibility into where tokens are being consumed
All stat cards now include descriptive tooltips explaining how each metric is calculated.
### One-Click Reprocessing for Knowledge Bases
Added a **Reprocess All** button to the knowledge base detail page. Reprocess every file in a knowledge base with a single click — each file restarts from the appropriate pipeline stage based on its current status.
* **Smart Routing**: Failed files retry from where they failed, completed files re-extract with the current schema
* **In-Progress Aware**: Files currently being processed are automatically skipped
* **Confirmation Dialog**: Shows exactly how many files will be reprocessed before you commit
* **Reliable Execution**: Reprocessing runs server-side, so you can navigate away without interrupting the operation
### Anthropic's Most Intelligent Model Now Available
Added support for **Claude Opus 4.6**, Anthropic's most intelligent model for building agents and coding.
**Specs:**
* 200K context window (1M available separately)
* 128K max output tokens
* Extended thinking support
Select `claude-opus-4-6` from the model dropdown in your agent configuration to start using it.
### Run Your Agents on a Schedule
* **Scheduled Tasks**: Create recurring tasks in Control Hub — pick an agent, write a prompt, set a schedule (presets or custom cron), and let it run automatically
* **Schedule Task Tool**: Enable `schedule_task` on any agent so it can create scheduled tasks during conversations — just ask it to "check stock prices every morning"
* **Execution History**: View past runs with status, timestamps, and links to the resulting conversations
* **Run Now**: Manually trigger any task on-demand
### Dictate Messages to Your Agents
* **Voice Input**: Click the microphone button in the chat input to dictate messages using your browser's built-in speech recognition
* **Hands-Free Convenience**: Speak naturally and your words are transcribed directly into the message box — edit before sending if needed
* **Works in Chrome**: Available on browsers that support the Web Speech API (Chrome, Edge). The button is automatically hidden on unsupported browsers.
### Start Conversations with Any Agent via Email
You can now email your AI agents directly and receive responses in your inbox. Conversations started via email appear in the dashboard like any other conversation.
**How it works:**
1. Every agent has an email address: `agent-{id}@updates.asteragents.com`
2. Send an email to start a new conversation
3. The agent processes your message (including attachments) and replies via email
4. Reply to continue the conversation — threading works automatically
**Features:**
* **Custom Email Aliases**: Set a friendly email slug (e.g., `my-sales-agent@updates.asteragents.com`) in agent settings
* **Attachment Support**: Send files as email attachments — they're uploaded and included in the agent's context
* **CC Preservation**: CC recipients are maintained throughout the email thread
* **Proper Email Threading**: Replies stay grouped in your email client
* **Secure by Default**: Only organization members can email agents — sender email is verified against your team
Find your agent's email address in the **Email Address** section of the agent editor in Control Hub.
### Automatically Run Agents When Files Are Uploaded
Knowledge bases can now trigger an AI agent conversation whenever a file finishes processing.
**How it works:**
1. Edit your knowledge base and enable "Trigger Agent on File Upload"
2. Select which agent should be triggered
3. Write instructions telling the agent what to do with the document
4. When any file completes processing, a new conversation is automatically started
**Use cases:**
* **Automated Document Analysis**: Summarize every document uploaded to your KB
* **Data Extraction Workflows**: Extract and report key information from invoices, contracts, or reports
* **Content Classification**: Have an agent categorize and tag incoming documents
* **Alert Generation**: Notify relevant parties when specific types of documents are uploaded
**Features:**
* Document content is automatically included in the agent's context
* Structured extraction data (if configured) is also provided
* View the triggered conversation directly from the file's menu
* Works with all upload methods: drag-and-drop, bulk upload, and email-to-KB
### Send Attachments Directly to Your Knowledge Bases via Email
* **Unique Email Addresses**: Each knowledge base gets its own email address (e.g., `kb-123@updates.asteragents.com`)
* **Just Forward or Send**: Email documents as attachments and they're automatically added to the knowledge base
* **Full Processing**: Files go through the same extraction, chunking, and embedding pipeline as manual uploads
* **Secure by Default**: Only organization members can add documents - sender email is verified against your team
**How it works:**
1. Find your KB's email address in the upload area of any knowledge base
2. Send an email with attachments to that address
3. Documents are automatically processed and added to your knowledge base
Perfect for quickly adding documents from your phone, forwarding emails with attachments, or integrating with systems that can send email.
### Let External Systems Sync Content to Your Knowledge Bases
* **Integration Sources**: Knowledge bases can now be marked as managed by an external integration (e.g., Salesforce, SharePoint)
* **Automatic Sync**: Integrations push content directly via API while the UI prevents manual uploads to avoid conflicts
* **Clear Indicators**: Managed KBs display a banner showing which integration controls them, plus a badge on KB cards
* **Retry Still Works**: Users can still retry failed file processing - only upload and delete are disabled
Perfect for enterprise teams that want to automatically sync documents from external systems while keeping knowledge base content consistent and controlled.
### Upload Presentations to Knowledge Bases and Conversations
* **PPTX Support**: Upload PowerPoint files (.pptx) to both knowledge bases and chat conversations
* **Full Content Extraction**: Agents automatically extract and understand slide text, speaker notes, and structure
* **Drag & Drop**: Works with the existing drag-and-drop interface in both KB uploads and chat
Add presentations to knowledge bases for searchable reference, or upload directly in chat for on-the-fly analysis.
### Find Anyone with AI-Powered People Search
* **New People Search Tool**: Your agents can now search for people by name, role, company, or any topic using Exa's specialized people search
* **Rich Profile Results**: Returns detailed information including profile URLs, titles, and relevant text snippets
* **Research Made Easy**: Perfect for sales prospecting, recruiting, competitive research, or finding experts in any field
Enable the tool in your agent's configuration under the Research category.
### Upload Outlook Emails to Your Knowledge Bases
* **MSG File Support**: Your knowledge bases now accept Outlook email files (.msg) - perfect for ingesting client correspondence, support tickets, and email threads
* **Full Email Extraction**: Automatically extracts sender, recipients, subject, body content, and attachment metadata from email files
* **Seamless Search**: Email content becomes fully searchable alongside your other knowledge base documents
Ideal for teams that need to make email history searchable and accessible to their agents.
### OpenAI's Flagship Model Now Available
Added support for **GPT-5.2**, OpenAI's flagship model for coding and agentic tasks.
**Specs:**
* 400K context window
* 128K max output tokens
* Reasoning token support
Select `gpt-5.2` from the model dropdown in your agent configuration to start using it.
### Interactive Charts for Agent Performance Insights
The Dashboard now includes a new **Analytics** tab with interactive visualizations to help you understand your agent usage patterns at a glance.
**Available Charts:**
* **Token Usage**: Stacked area chart showing input/output tokens over time with day or week grouping
* **Activity Trends**: Dual-axis line chart tracking conversations and messages over time
* **Top Users**: See which team members are using agents the most
* **Agent Performance**: Compare usage across your different agents
* **Tool Usage**: Breakdown of which tools are being called most frequently
* **Feedback Summary**: Donut chart showing helpful vs not helpful feedback with overall score
**How to Access:**
1. Go to **Dashboard**
2. Click **Analytics** in the top-right toggle (next to Conversations)
3. Use filters and date range to drill into specific data
All charts respect your existing filters (agent, user, date range, metadata) so you can analyze specific segments of your usage data.
### Visualize Agent Relationships with Interactive Graph View
The Agents tab in Control Hub now includes a new **Graph view** that visualizes how your agents connect to each other and their knowledge bases.
**Key Features:**
* **Interactive Graph**: See your agents and knowledge bases as nodes, with connections showing relationships
* **Relationship Types**: Solid arrows show agent-to-agent calls, dashed lines show knowledge base access
* **Click to Explore**: Click any node to see details in the info panel, double-click to navigate to the configuration page
* **Zoom & Pan**: Full control to explore complex agent networks with zoom controls and drag-to-pan
* **Legend**: Built-in legend explains what each node type and connection style means
**How to Use:**
1. Go to **Control Hub** → **Agents**
2. Click the **Graph** tab in the top-right toggle
3. Explore your agent ecosystem visually
Perfect for understanding complex multi-agent setups, identifying which agents share knowledge bases, and getting a bird's-eye view of your agent architecture.
### Use Built-in AI Capabilities Directly from Your Models
Your agents can now use powerful built-in tools from Anthropic, Google, and OpenAI—no external sandbox required.
**What's Available:**
* **Anthropic**: Code Execution, Computer Use, and Skills (Data Analysis, Stock Market, Think Tank, and more)
* **Google**: Code Execution for Gemini models
* **OpenAI**: Code Interpreter, Image Generation (DALL-E), and Web Search
**How to Enable:**
1. Go to **Control Hub** → **Agents** → select your agent
2. Provider tools appear automatically based on your selected model
3. Toggle on the tools you want
**Instant File Downloads:**
Files generated by provider tools (charts, images, exports) appear immediately during streaming—download buttons work right away without refreshing.
[Learn more about building agents →](/features/build-an-agent#provider-tools)
### See How Your Agents Think
* **Reasoning Tokens Visible**: When using models with extended thinking (like Claude with thinking enabled), you can now see the AI's reasoning process in real-time
* **Live Thinking Indicator**: Watch the sparkle icon pulse while your agent thinks, with a "Thinking..." label that updates to show how long reasoning took
* **Collapsible Display**: Reasoning automatically collapses when complete to keep your chat clean - click to expand and review the thought process anytime
* **Smart Defaults**: Reasoning stays open while streaming so you can follow along, then collapses to focus on the final response
Perfect for understanding complex decisions, debugging agent behavior, or simply seeing the "why" behind your agent's responses.
### Duplicate Configurations Instantly
**Clone Agents:**
* **Clone Any Agent**: New "Clone" action in the agent card menu creates a copy with all settings - name, model, tools, instructions, knowledge bases, and more
* **Quick Iteration**: Perfect for creating variations of existing agents without starting from scratch
* **Starts in Development**: Cloned agents default to development stage so you can test before releasing
**Clone Knowledge Bases:**
* **Clone Any Knowledge Base**: New "Clone" action copies embedding model, extraction model, and extraction schema settings
* **Starts Empty**: Cloned KBs are created without files so you can add fresh content
Find the clone option in the actions menu (⋮) on agent and knowledge base cards in Control Hub.
### Anthropic's Most Capable Model Now Available
* **Claude Opus 4.5**: Added support for Anthropic's flagship model, combining maximum capability with practical performance
* **Best For**: Complex specialized tasks, professional software engineering, and advanced agent workflows
* **Enhanced Capabilities**: Significant improvements in vision, coding, and agentic tasks
* **Easy Access**: Available now through Control Hub model settings
[Learn more from Anthropic →](https://www.anthropic.com/news/claude-opus-4-5)
### Organize Agents with Tags & Enhanced Agent Selection
**Agent Tags:**
* **Categorize Your Agents**: Add tags like "Sales", "Support", or "Analytics" to organize your agent library
* **Filter by Tag in Chat**: New sidebar in agent selector lets users browse agents by category
* **Instant Tag Creation**: Create new tags on-the-fly when editing agents - no setup required
* **Smart Discovery**: Tag counts show how many agents are in each category
**Improved Chat Experience:**
* **New Agent Selector Design**: Clearer "Switch Agent" button with chevron makes it obvious how to change agents
* **Better Visual Hierarchy**: Agent selection flow is more intuitive for first-time users
* **Responsive Layout**: Agent selector modal now properly displays on all screen sizes
Perfect for teams with many specialized agents - make it easy for everyone to find the right agent for their task.
[Learn about agent tags →](/features/build-an-agent#agent-tags)
### Give Your Team Agent Management Access Without Full Admin
New **Agent Manager** role lets team members create and manage agents without needing full admin privileges.
**What You Can Do:**
* AI engineers can build agents without accessing user management or billing
* Product managers can configure agent behavior without organization-wide admin access
* Admins keep control over sensitive settings like analytics and user invitations
[Learn how to set it up →](/settings/roles-permissions)
### Agents Can Now Cite Their Sources
* **Clickable Citations**: When agents reference knowledge base files, they now include clickable citation chips showing the document name and page numbers
* **Instant Document Viewing**: Click any citation to open the full document, automatically jumping to the cited page
* **Verifiable Sources**: See exactly which documents and pages your agent used to answer your questions
Perfect for research, compliance, and any workflow where you need to verify information sources.
### Find Past Conversations Instantly
* **Search Button in Sidebar**: New search button appears at the top of your conversations list in the sidebar - just click to start searching
* **Search Your Chat History**: Quickly find any past conversation by searching through what you and your agents discussed
* **Fast and Accurate**: Get instant results as you type - no more scrolling through endless chat history
* **Search Within Messages**: Finds conversations based on the actual content of your messages and agent responses, not just titles
Perfect for recalling that data analysis your agent ran last week, finding specific recommendations from previous chats, or referencing important decisions made in past conversations with your team's agents.
Fixed `RangeError: Invalid time value` errors in dashboard endpoints when invalid or missing date parameters were provided.
### New RESTful API Endpoints for User Management
We've modernized our user and invitation management APIs with a comprehensive refactor that improves clarity, follows REST conventions, and provides better separation of concerns.
#### New Endpoints
**User Management:**
* [GET /admin/users](/api-reference/users/list-organization-users) - List active organization users with pagination and filtering
* [DELETE /admin/users](/api-reference/users/remove-users-from-organization) - Remove multiple users in a single request
**Invitation Management:**
* [GET /admin/invitations](/api-reference/invitations/list-organization-invitations) - List invitations with status filtering (pending, accepted, revoked)
* [POST /admin/invitations](/api-reference/invitations/create-bulk-invitations) - Send bulk invitations with smart handling for existing users
* [DELETE /admin/invitations](/api-reference/invitations/revoke-invitations) - Revoke pending invitations
#### What's Better
* **Separation of Concerns**: Users and invitations are now managed through separate, focused endpoints instead of one combined endpoint
* **Proper REST Conventions**: Uses appropriate HTTP methods (GET, POST, DELETE) instead of POST for everything
* **Better Pagination**: All list endpoints support `limit` and `offset` parameters for efficient data retrieval
* **Status Filtering**: Filter invitations by status (pending, accepted, revoked, expired) to find exactly what you need
* **Consistent Response Format**: All endpoints use camelCase and return standardized error responses
#### Migration Guide
**No immediate action required!** All old endpoints remain fully functional during the transition period.
When you're ready to migrate, here are the mappings:
**Old → New:**
* `POST /admin/bulkInvite` → `POST /admin/invitations`
* `POST /admin/bulkRemove` → `DELETE /admin/users` (for members) or `DELETE /admin/invitations` (for pending invitations)
* `GET /admin/getUsersInOrg` → Make two requests: `GET /admin/users` + `GET /admin/invitations` and merge results
**Response Field Changes:**
* `user_id` → `userId`
* `invitation_id` → `invitationId`
* `membership_id` → `membershipId`
* `expires_at` → removed (use `status` field instead)
#### Backward Compatibility
* ✅ All legacy endpoints continue to work identically
* ✅ No breaking changes to existing integrations
View the [legacy endpoint documentation](/api-reference/endpoint/bulk-invite) for migration details or explore the [new Admin API documentation](/api-reference/introduction) to get started.
### Enhanced Multi-Organization Support
* **Organization-Scoped Metadata**: User metadata is now isolated per organization membership
* **Architecture Update**: Metadata moved from global `user.publicMetadata` to organization-scoped `userOrganization.metadata`
* **Multi-Org Improvement**: Users who belong to multiple organizations now maintain separate metadata for each organization
* **Backward Compatible**: All existing API responses maintain the same structure - no breaking changes
* **New Endpoint**: Added [Update User Metadata](/api-reference/endpoint/update-user-metadata) API documentation
This enhancement ensures that when users belong to multiple organizations, each organization can maintain independent metadata (roles, departments, custom fields) without affecting the user's profile in other organizations.
### Latest Anthropic Model Now Available
* **Claude Sonnet 4.5**: Added support for Anthropic's newest model with improvements in coding, agentic workflows, and reasoning capabilities
* **Learn More**: Read the [official announcement](https://www.anthropic.com/news/claude-sonnet-4-5) from Anthropic
* **Easy Access**: Available immediately for all agents through Control Hub model settings
Keep your agents powered by the latest AI technology as new models are released.
### Share Direct Links to Agents and Pre-filled Questions
* **Agent Deep Linking**: Share direct links to specific agents using `?agent=` URL parameter - supports both agent ID and name matching
* **Question Pre-filling**: Pre-fill the chat input with `?q=` parameter for streamlined user workflows and support scenarios
* **Smart URL Handling**: Question parameters are automatically cleaned from the URL after use to prevent re-application on navigation
* **New Chat Only**: Question pre-filling only works for new conversations, preserving existing chat history
Perfect for creating shareable support links, onboarding flows, or quick access to specific agent interactions.
### Detailed Token Usage Breakdown
* **Input/Output Token Tracking**: Dashboard now shows breakdown of token usage with separate counts for input tokens (user messages and system prompts) and output tokens (AI responses)
* **Cleaner Analytics Display**: Token usage cards now display the breakdown in a clean "X in / Y out" format alongside the total token count
* **Better Cost Insights**: Understanding input vs output token usage helps organizations better understand their AI costs, as many providers charge different rates for input and output tokens
Perfect for organizations wanting deeper insights into their AI usage patterns and associated costs.
### Enhanced JSON Schema Configuration Experience
* **Larger Modal Windows**: Expanded schema editor and viewer modals for better editing experience with complex schemas
* **Real-time Validation**: Live JSON validation with required field checking as you type
* **Copy Example Button**: Quickly insert schema templates to get started with structured data extraction
* **Better Form Layout**: Improved spacing and helpful subtext for all schema configuration fields
* **Fixed Scrolling**: Proper internal scrolling for long schemas without modal overflow issues
Makes configuring structured data extraction from your documents much easier and more intuitive.
### Persistent Python Execution Environment
* **Thread-based State Management**: Python sandboxes now maintain variables, imports, and memory state across multiple code executions within the same conversation
* **5-6x Performance Improvement**: Subsequent Python executions are dramatically faster by reusing existing sandbox environments
* **Smart File Detection**: Automatically detects when files haven't changed to skip redundant uploads
* **Seamless Experience**: No changes needed - agents automatically benefit from persistent state when running Python code
### Enhanced File Processing
* **Larger File Support**: Increased file upload limits for knowledge bases and conversations
* **Improved Timeouts**: Extended processing time limits (up to 800s) for large document processing
* **Better Error Messages**: More helpful feedback when file processing encounters issues
* **Token Limit Management**: Intelligent handling of large files to stay within AI model token constraints
Perfect for data analysis workflows that build on previous results and processing large document collections.
### Advanced Web Research with AI-Powered Analysis
* **Deep Research Capability**: New Perplexity-powered tool for comprehensive research on any topic with multi-step analysis
* **Citation-Rich Results**: Automatically includes source citations and references for all research findings
* **Advanced Query Processing**: Handles complex research requests with follow-up questions and detailed investigation
* **Token Usage Tracking**: Full transparency on research costs with detailed token consumption reporting
Transform your agents into powerful research assistants capable of conducting thorough, well-sourced investigations on any topic.
### User Re-invitation Support
* **Fixed Bulk Invite**: Previously removed users can now be re-added without "account already exists" errors
* **Smart Detection**: Automatically handles new users (invitation) vs existing users (direct membership)
### Review and Edit AI-Extracted Data
* **File Review Interface**: New side-by-side interface for reviewing knowledge base files with extracted structured data alongside the original document
* **Interactive Data Editor**: Schema-aware editor for modifying AI-extracted structured data with rich field types (text, numbers, dates, objects, arrays)
* **Document Preview**: Multi-format document viewer supporting PDF, Office documents, images, and text files with presigned URL security
* **Real-time Editing**: Live editing with change tracking and save functionality - perfect for correcting or enhancing AI extractions
* **Navigation Integration**: Access the review interface directly from knowledge base file lists via "Review Extracted Data" menu option
Transform your knowledge base workflow by easily reviewing and refining the structured data extracted from your documents, ensuring maximum accuracy for agent searches.
### Rename Conversations
* **Rename in Sidebar**: Quickly rename any conversation from the chat sidebar's "More" menu
* **Clear Names**: Remove a custom name to revert to the automatic preview title
* **API**: PATCH `/api/renameConversation` with `{ threadId, name }`; only conversation owners or org admins can rename within their organization
### New OpenAI Models
* **GPT-5 and GPT-5-mini**: Now available as chat models across agents and API
* **Temperature Handling**: Defaults applied per provider requirements for GPT-5 family
### Advanced Knowledge Base Capabilities
* **Python Sandbox Integration**: New way to pull knowledge base content directly into the Python sandbox environment for advanced document analysis and processing workflows
* **Structured Data Extraction**: Added optional extraction schema configuration to knowledge bases, enabling AI-powered extraction of structured information from documents using custom JSON schemas
* **Snowflake Personal Access Tokens**: Added support for Personal Access Tokens (PATs) for more secure and flexible Snowflake database authentication
### Expanded File Format Support
* **HTML File Support**: Upload .htm and .html files directly to knowledge bases and conversations for web content analysis
* **Enhanced CSV Processing**: Improved CSV support with better table chunking strategies for more accurate analysis of tabular data
### Chat & Agent Experience Improvements
* **Attachment-Only Messages**: Send messages with just file attachments without requiring text - perfect for quick file sharing with agents
* **Prompt Variables Display**: Agent configuration screen now shows all supported prompt variables, making it easier to customize agent behavior
* **Enhanced Chat Sidebar**: Message previews and improved formatting in the conversation sidebar for better navigation
* **Grok 4 Model**: Added support for xAI's latest Grok 4 model with enhanced reasoning capabilities
### CSV and PowerPoint Support for Your Knowledge Bases
* **CSV File Upload**: Upload spreadsheets and data files directly to your knowledge bases - agents can now search and analyze your tabular data seamlessly
* **PowerPoint Support**: Add your presentation files (.pptx) to knowledge bases so agents can reference slide content, charts, and presentation materials
* **Enhanced File Processing**: Improved how agents understand and chunk table data for better search results and more accurate responses
Now your knowledge bases support an even wider range of business documents, making it easier to give your agents access to all your important information in whatever format you have it.
### Latest AI Model from xAI
* **Grok 4 Support**: Your agents now have access to xAI's latest and most advanced Grok 4 model, offering enhanced reasoning capabilities and improved response quality
* **Expanded Model Options**: Choose from an even broader selection of cutting-edge AI models to power your agents based on your specific needs
Stay at the forefront of AI with access to the newest models as soon as they become available.
### Upload Documents and Give Your Agents Instant Access to Your Knowledge
* **Knowledge Bases**: Create organized collections of documents that your agents can search and reference during conversations
* **Multi-Format Support**: Upload PDF, DOCX, TXT, and Markdown files - agents automatically extract and understand the content
* **Semantic Search**: Agents can search your knowledge bases using natural language queries to find relevant information instantly
* **Agent Integration**: Configure which agents have access to which knowledge bases using the new `search_knowledge_base` tool
* **Smart Processing**: Files are automatically chunked and embedded for optimal search performance with real-time processing status
* **Enterprise Ready**: All knowledge bases are organization-scoped with admin controls for secure document management
Transform your documents into searchable knowledge that your agents can access instantly. Perfect for company policies, research papers, technical documentation, and any reference materials your team needs.
### Tag and Organize Your Team with Custom Metadata
* **Custom User Metadata**: Organization admins can now add custom key-value metadata to any user (e.g., department: "Sales", role: "Manager", team: "West Coast")
* **Flexible Metadata Editor**: Add, edit, and remove metadata fields with an intuitive interface that supports any custom properties your organization needs
* **Dashboard Filtering**: Use custom metadata to filter conversations and analytics in the Dashboard - perfect for analyzing usage by department, role, or team
* **Visual Metadata Display**: User metadata appears as badges throughout the interface, making it easy to identify user attributes at a glance
Ideal for organizations that need to categorize and analyze user activity by custom attributes beyond basic user profiles.
### SQL Query Results as Downloadable Files
* **CSV Downloads**: SQL query results from Snowflake and other databases now include convenient "Download CSV" buttons
* **Automatic Attachments**: Query results are automatically saved as downloadable attachments, making it easy to export data for further analysis
* **Enhanced File Handling**: Generated files from agent workflows are now properly tracked and made available throughout conversations
Perfect for data analysts who need to quickly export query results or share findings with their team.
### Create Professional Presentations with AI
* **Generate PowerPoint Tool**: Your agents can now create complete PowerPoint presentations from structured data - no templates required!
* **Rich Content Support**: Include text, bullet lists, tables, charts, and images across multiple slide layouts
* **Professional Layouts**: Choose from built-in PowerPoint layouts including title slides, content slides, and section headers
* **Instant Downloads**: Generated presentations are automatically saved and available for immediate download as .pptx files
* **Smart Formatting**: Agents understand how to structure slide content logically with proper hierarchy and visual organization
Transform your data analysis, reports, and summaries into polished presentations instantly. Just ask your agent to "create a PowerPoint about..." and watch them build comprehensive slide decks tailored to your needs.
### Execute Python Tool with File Access
* **File Upload Support**: Execute Python now automatically uploads conversation attachments to the sandbox at `/mnt/data/` paths
* **Enhanced Output Display**: Improved categorization of stdout, stderr, and results with better visual formatting
* **Better Error Handling**: More detailed error messages and tracebacks for debugging
Your Python code can now directly access uploaded files from conversations, making data analysis workflows much more seamless.
### Enhanced Control Hub & Model Management
* **Reorganized Model Settings**: Split model configuration into logical sections:
* **"Providers"** tab for API key management (OpenAI, Azure, etc.)
* **"Models"** tab for selecting default chat and embedding models
* **Organization-Wide Defaults**: Set default models for your entire organization that all agents inherit
* **Azure OpenAI Improvements**: Better handling of Azure model deployments and instances
* **Model Instance Management**: Clean interface for managing Azure deployment names and configurations
* **Enhanced Embedding Model Support**: Organizations can now configure default embedding models from either OpenAI or Azure OpenAI
* Supports standard OpenAI embedding models (text-embedding-3-small, text-embedding-3-large, etc.)
* Full support for Azure OpenAI embedding deployments with automatic deployment name resolution
* Intelligent model routing: automatically uses the correct provider based on your organization's configuration
* Backward compatible: existing setups continue to work without any changes
### Chat Experience Enhancements
* **Beautiful Typing Indicator**: New animated typing indicator shows when agents are thinking with smooth dot animations
* **Improved Agent Selection**: Enhanced agent selector on chat page with better visual hierarchy
* **Better Spacing & Layout**: Refined chat interface with improved message spacing and visual flow
* **Attachment Visibility**: Attachments now properly display when viewing conversation history
* **Enhanced Error Handling**: Better error messages and handling for old chat formats
### Connect Your Agents to Remote MCP Servers
* **Remote MCP Server Connections**: Your agents can now connect to MCP servers hosted anywhere on the internet, not just locally installed ones
* **Easy Server Setup**:
* Choose from popular pre-configured servers (Slack, GitHub, Neon Postgres, Filesystem)
* Or configure custom remote MCP servers with just a URL and optional headers
* Test connections before saving to ensure everything works
* **Expanded Tool Library**: Connect to remote MCP services to give your agents access to:
* External APIs and databases
* Third-party services and platforms
* Custom tools hosted by your team or organization
* **Seamless Integration**: Once configured, remote MCP tools work exactly like built-in tools - your agents automatically discover and use them during conversations
This opens up a whole new world of possibilities for extending your agents with external capabilities, whether from public MCP servers or your own custom implementations.
### Upload and Analyze Word Documents
* **DOCX File Support**: Your agents can now read and analyze Microsoft Word documents (.docx files) uploaded to conversations
* **Smart Document Processing**: Agents automatically extract text, preserve heading structure, and convert tables to readable markdown format
* **Comprehensive Understanding**: Ask questions about document content, get summaries, or have agents analyze the structure and formatting of your Word documents
* **Seamless Integration**: Simply drag and drop .docx files directly into chat - agents will automatically process and understand the content
Perfect for analyzing reports, proposals, contracts, and other business documents with your AI agents.
### Drag & Drop File Uploads
* **Intuitive File Uploads**: Simply drag and drop files directly into the chat interface - no more clicking upload buttons
* **Visual Feedback**: See helpful drag overlays that show supported file types and upload status
* **Real-time Progress**: Watch upload progress bars as your files are processed and made available to your agents
* **Multi-file Support**: Drop multiple files at once - supports images, documents, spreadsheets, and more
* **Paste Images**: Copy and paste images directly from your clipboard into conversations
Makes sharing files with your agents as natural as dropping them into the conversation.
### Latest AI Models & Enhanced Search
* **New Claude Models**: Access to Anthropic's latest Claude Sonnet 4 and Opus 4 models (May 2025) with significantly improved reasoning capabilities
* **Enhanced Search API**: Switched to Brave Search API for faster, more reliable web search results with better privacy
* **Expanded Model Selection**: Added support for new OpenAI models including GPT-4.1 and o4-mini for specialized reasoning tasks
* **Performance Optimization**: Removed deprecated models to streamline the model selection experience
Your agents now have access to cutting-edge AI models with enhanced reasoning, better search capabilities, and improved response quality.
### Multi-Agent: Configurable Call Agent
* Introduced a user interface for configuring the `call_agent` tool, enabling users to select which other agents a specific agent is permitted to call.
* The system prompt for an agent now dynamically includes the ID, name, and description of its designated callable agents, providing necessary context for the `call_agent` tool.
* The agent configuration page has been updated to list available agents, allowing for easy selection when setting up the `call_agent` functionality.
### Agent Stage
* Agents can now be designated as either `Development` or `Released`.
* Agents in `Development` are now visually grouped separately at the bottom of agent selection dropdowns (Chat page, Dashboard).
* Share your Conversation: Agent conversation links can now be shared with anyone in your organization
* Fix bug where tool calls did not appear in the proper order in an agent conversation
### File Uploads & Processing
* Support adding Excel files into agent conversations
* Refactored file uploads to use presigned URLs for direct R2 storage uploads, improving reliability and adding progress tracking.
* Implemented asynchronous content pre-extraction for non-image files (e.g., PDF text) upon upload.
### UI Enhancements
* Added visual progress indicators for file uploads.
* Improved handling of pasted images in chat input.
* Updated chat display to use icons for non-image attachments.
* Users can now delete their own conversations from the chat list.
* New user\_manager role, for users that should be able to manage the other users in their organization but not change AI Settings
* Fixed an issue where the dashboard date range picker could become undefined
* Added support for OpenAI's `o3-mini` model (Jan 31).
* Allowed line breaks in user chat messages for better readability (Jan 6).
* Improved URL scraping to correctly extract links found within HTML table cells (Jan 3).
### Pinecone Integration
* Added a new integration with Pinecone for semantic search
* This integration is now available in the Control Hub under the Integrations section
### Bug Fixes
* Errors encountered when chatting with agents should now be visible on the chat page
* Fixed a bug where message pruning based on token count was not working
* Fixed a bug where execute python was not working
* Fixed a bug where belonging to more than 10 organizations was causing failed requests
### Filter and Search Postgres
* Added a new tool to filter and search Postgres embeddings for a given query in a specified table
* This allows your agent to filter the table with a WHERE clause and then semantically search ONLY the filtered results
* This tool is now available in the agent configuration under the Tools section
### TikTok Tool
* Added new tool to search TikTok for videos matching a search term or hashtag
* This tool is now available in the agent configuration under the Tools section
### Reddit Tools
* Added new tools to search and read Reddit
* Search Reddit for posts matching a search term, across all subreddits
* Read posts from a specified subreddit, with optional filtering for time period and post type
* This tool is now available in the agent configuration under the Tools section
### PDF Understanding
* Agents can now understand PDFs that are uploaded in a conversation
* PDF Understanding is powered by two features:
* Traditional OCR to extract readable text from the PDF and provide it to the Agent
* PDF-to-image conversion that enables the Agent to "see" the first 5 pages of the PDF
PDF Understanding is currently in beta and has multiple limitations
Limitations:
1. Only the first 5 pages of the PDF are processed into images that the Agent can "see"; only the readable text from remaining pages is provided to the Agent
2. PDFs bigger than 3MB may fail to process
3. PDF-to-image conversion causes a delay in the Agent's response time
4. PDF URLs that are accessed via the Scrape URL tool do not have PDF-to-image conversion enabled, so the agent will only get extracted text
### Copy Code to Clipboard
* Code blocks now have a copy button to easily copy the code to your clipboard
### Terms of Service and Privacy Policy
* Users are now required to accept the Terms of Service and Privacy Policy to sign up for Aster Agents
### Agent Feedback on Dashboard
* Added feedback score and total feedback given to your agents on the Dashboard
* Feedback for specific conversations will be shown in the conversation list. Hover to see the users who provided feedback.
### API Keys!
* API Keys are now accessible from the Control Hub Settings page
* API Keys are valid for 6 months
* You can use your API key to communicate with your agents via the API
### Perplexity API Citations Support
* Added support for Perplexity's new citations feature
* Citations are now automatically included in `ask_web` responses
* Each citation is displayed as a clickable source link below the response
This update integrates with Perplexity's [public release of citations](https://docs.perplexity.ai/changelog/changelog#citations-public-release-and-increased-default-rate-limits) and helps provide more transparency about where information is sourced from.
### Agent Feedback
* Added the ability to provide feedback on your conversations once the agent has responded at least once
* Let us know if an agent's responses were helpful with a simple thumbs up or down
* Changed your mind? You can always update your feedback later
Soon, your feedback will be automatically shared with your AI Admins via the Aster Agents Dashboard.
### Improvements to Agent Configuration
* Group tools by required integrations on the agent configuration page
* If an integration is not setup, its tools cannot be turned on for the agent
### New Model Provider: xAI
* Added xAI as a supported LLM provider for building agents
In order to connect your xAI account, all you need is:
```plaintext theme={null}
XAI_API_KEY: "Your xAI API key"
```
### New Tool: Call API
* Introduced a new tool to let your agent call API endpoints directly
* Supports GET, POST, PUT, and DELETE request types
This tool is now available in the agent configuration under the Tools section.
### Multi-organization support
* Added the ability to create multiple organizations and switch between them
* Each organization has its own set of agents, conversations, and users
* Admins can invite new members to their organization via email
### New Model Provider: Google AI
* Added Google AI as a supported LLM provider for building agents
In order to connect your Google AI account, all you need is:
```plaintext theme={null}
GOOGLE_API_KEY: "Your Google AI API key"
```