mcp-bridge

This configuration launches an MCP server in Docker that internally calls the llama.cpp OpenAI-compatible API.

Startup

docker compose up -d --build

Related MCP server: openai-tool2mcp

Shutdown

docker compose down

Connection Details

• MCP endpoint: http://localhost:8000/mcp

• Transport: streamable-http

Connection Verification

Even if a 406 is returned, it is normal as a reachability check. This is because the MCP endpoint requires an Accept header.

curl -i http://localhost:8000/mcp

Claude Code Configuration

Claude Code's MCP is managed in ~/.claude.json.

If you are configuring it manually for a local scope (specific project only), please apply the format of client-config.example.json to ~/.claude.json.

Recommended: Add via CLI

claude mcp add --transport http local-agent-helper http://localhost:8000/mcp

Verify Configuration

claude mcp list
claude mcp get local-agent-helper

Notes

• Adding --scope user makes it available for all projects.

• Using --scope project saves it to .mcp.json directly under the project.

Environment Variables

You can copy .env.example to .env to use it.

• MCP_TRANSPORT (Default: streamable-http)

• FASTMCP_HOST (Default: 0.0.0.0)

• FASTMCP_PORT (Default: 8000)

• LLAMA_CPP_SERVER_URL

• LLAMA_CPP_SERVER_MODEL

• LLAMA_CPP_SERVER_TIMEOUT