Skip to main content

MCP Server

The ABI MCP (Model Context Protocol) server exposes every loaded ABI agent as an MCP tool. This lets Claude Desktop, Cursor, VS Code, and any MCP-compatible client invoke ABI agents directly.

Starting the MCP server​

# STDIO mode (for Claude Desktop)
make mcp
# or
uv run python -m naas_abi.apps.mcp.mcp

# HTTP mode (for remote / multi-client)
make mcp-http
# or
uv run python -m naas_abi.apps.mcp.mcp --transport http --port 8000

Claude Desktop integration​

Add to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

{
  "mcpServers": {
    "abi": {
      "command": "uv",
      "args": ["run", "python", "-m", "naas_abi.apps.mcp.mcp"],
      "cwd": "/path/to/your/abi",
      "env": {
        "OPENROUTER_API_KEY": "sk-or-...",
        "ABI_API_KEY": "your-api-key"
      }
    }
  }
}

Restart Claude Desktop. ABI agents will appear as available tools.

How it works​

At startup, the MCP server:

  1. Starts the ABI Engine and loads all configured modules.
  2. Discovers every registered agent.
  3. Registers one MCP tool per agent, using the agent's description as the tool description.

Tool calls are routed directly to the corresponding agent's run() method. No additional configuration is needed when agents are added to or removed from the engine.

See [[adr/20250730_mcp-server-dynamic-agent-discovery|ADR: MCP Dynamic Agent Discovery]].

Testing the MCP server​

make mcp-test

This runs MCP protocol validation tests to confirm all tools are registered and respond correctly.