rag-forge serve

Start the RAG-Forge server

Synopsis


rag-forge serve --mcp [options]

Description

serve launches the RAG-Forge MCP (Model Context Protocol) server, which exposes your RAG pipeline as a set of tools that AI assistants — such as Claude Desktop or any MCP-compatible client — can call directly.

The --mcp flag is required. Without it the command exits with an error.

By default the server starts on stdio transport, which is the correct mode for local AI assistant integration (e.g., Claude Desktop config). Pass --transport http to start an SSE-based HTTP server instead, which is useful for remote clients, n8n workflows, or browser-based tooling. The HTTP server listens on http://localhost:<port>/sse.

⚠️ Security. The HTTP transport is unauthenticated by default and has no TLS, rate-limiting, or origin checking. Do not expose it directly to the public internet. For any networked deployment, place it behind a reverse proxy (nginx, Caddy, Cloudflare Tunnel) that terminates TLS, enforces authentication, and restricts the allowed clients. For single-machine local use, bind it only to 127.0.0.1.

The MCP server binary must be built before running this command. If it has not been built yet, run pnpm run build from the repository root first.

Options

Flag	Default	Description
`--mcp`	—	Launch the MCP server (required)
`--transport <type>`	`stdio`	Transport mode: `stdio` \| `http`
`-p, --port <number>`	`3100`	Port number for HTTP transport

Examples

Start the MCP server on stdio (for Claude Desktop)


rag-forge serve --mcp

Start the MCP server on HTTP for remote clients


rag-forge serve --mcp --transport http

Start the MCP server on a custom HTTP port


rag-forge serve --mcp --transport http --port 4000

MCP Overview — how to connect AI assistants to the RAG-Forge MCP server