Skip to main content

Overview

The Model Context Protocol (MCP) is an open standard that lets AI agents interact with external tools and services through a unified interface. Aster Agents supports connecting to any remote MCP server, giving your agents access to thousands of third-party tools without custom integration work. When an agent has MCP servers assigned, it automatically discovers and can call all tools provided by those servers during conversations.

Features

  • Any Remote Server: Connect to any MCP server hosted on the internet by URL
  • Two Transports: Streamable HTTP (recommended) and Server-Sent Events (SSE)
  • Auth Headers: Pass API keys or Bearer tokens via custom headers
  • Auto-Discovery: Agent automatically discovers available tools at chat time
  • Multi-Server: Assign multiple MCP servers to a single agent
  • Organization-Level: Configure servers once, assign to any agent

Prerequisites

  • An MCP server URL (hosted by a third party or self-hosted)
  • Any required authentication credentials (API key, Bearer token, etc.)
  • Admin access to your Aster Agents organization

Setup Guide

1

Add an MCP Server

Navigate to Control Hub > Integrations and scroll down to MCP Servers. Click Add Server.
2

Configure the Server

Fill in the server details:
  • Name: A display name (e.g. “Hugging Face”, “Cloudflare Docs”)
  • Transport: Choose Streamable HTTP (recommended) or SSE
  • URL: The server’s MCP endpoint URL
  • Headers (optional): Add authentication headers (e.g. Authorization: Bearer your-api-key)
3

Test the Connection

After adding the server, click the play button on the server card to test connectivity. A successful test will report how many tools were discovered.
4

Assign to an Agent

Go to Control Hub > Agents, edit an agent, and scroll to the MCP Servers section. Check the servers you want this agent to use.

Transport Types

TransportBest ForHow It Works
Streamable HTTPProduction (recommended)Uses HTTP POST for requests and optional SSE for streaming responses. Most reliable for serverless environments.
SSELegacy serversUses Server-Sent Events for bidirectional communication. Supported by older MCP server implementations.
If the server documentation doesn’t specify, try Streamable HTTP first.

Example: Public MCP Servers

Many services now offer hosted MCP servers you can connect directly:
ServerURLTransportAuth
Hugging Facehttps://hf.co/mcpHTTPNone
Cloudflare Docshttps://docs.mcp.cloudflare.com/sseSSENone
Check awesome-remote-mcp-servers on GitHub for a community-maintained list of public MCP servers.

How It Works at Runtime

When a user sends a message to an agent with MCP servers:
  1. Connect — Aster opens a connection to each assigned MCP server
  2. Discover — Tools are fetched from the server via the MCP protocol
  3. Merge — MCP tools are combined with the agent’s built-in tools
  4. Execute — The AI model can call any MCP tool during the conversation
  5. Cleanup — Connections are closed after the response completes
MCP tools appear alongside built-in tools from the model’s perspective. The agent decides which tools to use based on the conversation context.

Security

  • Headers are encrypted — Auth headers are stored securely and redacted ([REDACTED]) in API responses
  • Organization-scoped — Only admins can add or modify MCP servers
  • Per-agent assignment — Each agent only accesses the servers explicitly assigned to it
  • No outbound data — Aster only sends tool call parameters to MCP servers, never conversation history

Troubleshooting

  • Verify the URL is correct and the server is running
  • Check that you selected the right transport type (HTTP vs SSE)
  • Ensure any required auth headers are included
  • Try the URL in a browser to confirm it’s accessible
  • The server may be running but has no tools registered
  • Check the server’s documentation for setup requirements
  • Some servers require authentication before exposing tools
  • Check that auth credentials haven’t expired
  • Verify the server supports the parameters the model is sending
  • Review the conversation for error messages in tool results
  • Confirm the MCP server is assigned to the agent in Control Hub
  • Ensure the server is enabled (toggle on the server card)
  • Try prompting the agent to explicitly use a specific tool by name