elizaos--eliza/packages/examples/mcp

elizaOS MCP Agent Server Examples

This directory contains MCP (Model Context Protocol) server implementations that expose an elizaOS agent as an MCP server. This allows any MCP-compatible client (like Claude Desktop, VS Code, etc.) to interact with your AI agent.

Uses real elizaOS runtime with OpenAI and SQL plugins!

Available Examples

Framework	Language	Directory
@modelcontextprotocol/sdk	TypeScript	.

What is MCP?

The Model Context Protocol (MCP) is an open protocol that standardizes how AI applications communicate with external tools and data sources. By exposing your elizaOS agent as an MCP server, any MCP client can:

• Send messages to your agent
• Receive responses
• Access agent metadata and capabilities

Common MCP Tools

All implementations expose the same tools:

chat

Send a message to the agent and receive a response.

Input Schema:

{
  "type": "object",
  "properties": {
    "message": {
      "type": "string",
      "description": "The message to send to the agent"
    },
    "userId": {
      "type": "string",
      "description": "Optional user identifier"
    }
  },
  "required": ["message"]
}

Returns: Agent's response text

get_agent_info

Get information about the agent.

Returns: Agent name, bio, and capabilities

Quick Start

cd packages/examples/mcp
bun install
OPENAI_API_KEY=your-key bun run start

Configuration

All examples require an OpenAI API key:

export OPENAI_API_KEY=your-key

Optional configuration:

• MCP_PORT - Port for HTTP transport (default: 3000)
• OPENAI_BASE_URL - Custom OpenAI-compatible endpoint
• OPENAI_SMALL_MODEL - Model for quick responses
• OPENAI_LARGE_MODEL - Model for complex responses

Connecting to Claude Desktop

Add to your Claude Desktop config (claude_desktop_config.json):

{
  "mcpServers": {
    "eliza": {
      "command": "bun",
      "args": ["run", "start"],
      "cwd": "/path/to/eliza/examples/mcp",
      "env": {
        "OPENAI_API_KEY": "your-key"
      }
    }
  }
}

Testing

cd packages/examples/mcp
bun run test