MCP Adapter¶

The MCP adapter exposes ToolRegistry tools via the Model Context Protocol for LLM integration.

Overview¶

The adapter:

• Registers list_tools and call_tool MCP handlers
• Supports multiple transport mechanisms (stdio, SSE, Streamable HTTP)
• Reads from RouteTable at request time for real-time sync
• Handles async and sync tools transparently
• Serializes results to JSON-compatible strings

Quick Start¶

Streamable HTTP Transport (Recommended)¶

from toolregistry import ToolRegistry
from toolregistry_server import RouteTable
from toolregistry_server.mcp import create_mcp_server, run_streamable_http
import asyncio

registry = ToolRegistry()

@registry.register
def greet(name: str) -> str:
    """Greet someone by name."""
    return f"Hello, {name}!"

route_table = RouteTable(registry)
server = create_mcp_server(route_table)

asyncio.run(run_streamable_http(server, host="0.0.0.0", port=8000))

stdio Transport¶

For subprocess-based communication (used by Claude Desktop, etc.):

from toolregistry_server.mcp import create_mcp_server, run_stdio
import asyncio

server = create_mcp_server(route_table)
asyncio.run(run_stdio(server))

SSE Transport¶

For Server-Sent Events over HTTP:

from toolregistry_server.mcp import create_mcp_server, run_sse
import asyncio

server = create_mcp_server(route_table)
asyncio.run(run_sse(server, host="0.0.0.0", port=8000))

Transport Comparison¶

MCP Client Configuration¶

Claude Desktop¶

Add to your Claude Desktop configuration:

{
  "mcpServers": {
    "my-tools": {
      "command": "toolregistry-server",
      "args": ["mcp", "--config", "config.json"]
    }
  }
}

HTTP-based Clients¶

Connect to the server URL:

http://localhost:8000/mcp

Error Handling¶

Tool errors are returned as structured MCP error responses with appropriate error codes.

API Reference¶

See the MCP API Reference for detailed documentation.