@rag-forge/mcp

Model Context Protocol server exposing RAG-Forge pipeline operations as agent-callable tools.

@rag-forge/mcp is the MCP (Model Context Protocol) server for RAG-Forge. It lets agents — Claude Desktop, Claude Code, Copilot CLI, Cursor, or any MCP-compatible client — query a RAG pipeline, run audits, score against the RAG Maturity Model, and inspect indexed chunks without leaving the conversation.

Install

npm install -g @rag-forge/mcp
# or
pnpm add -g @rag-forge/mcp

You also need the Python runtime that powers the actual pipeline:

uv pip install rag-forge-core rag-forge-evaluator rag-forge-observability

Run as an MCP server

stdio transport (Claude Desktop, Cursor)

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

{
  "mcpServers": {
    "rag-forge": {
      "command": "rag-forge-mcp",
      "args": ["--stdio"]
    }
  }
}

HTTP transport (web agents, MCP gateways)

rag-forge-mcp --port 3100

Then point your MCP client at http://localhost:3100/mcp.

From the main CLI

If you have @rag-forge/cli installed, you can also start the MCP server through it:

rag-forge serve --mcp --stdio
# or
rag-forge serve --mcp --port 3100

What the agent can do

Once connected, the agent gains these tools:

| Tool | Purpose | |---|---| | rag_forge_query | Run a RAG query against the indexed corpus and return retrieved chunks + generated answer | | rag_forge_audit | Run an evaluation audit against telemetry or a golden set, returning RMM level + metric scores | | rag_forge_cost | Estimate the cost of a planned audit before running it | | rag_forge_inspect | Inspect indexed chunks, embedding stats, and retrieval results for a query | | rag_forge_drift_report | Compare current pipeline state against a saved baseline | | rag_forge_golden_add | Add a new question/answer pair to the golden set | | rag_forge_assess | One-shot RMM assessment without running a full audit |

All tools return structured JSON the agent can reason about — perfect for agentic debugging loops where the model decides which audit to run, interprets the report, and proposes the next experiment.

Project context

Run from inside a RAG-Forge project directory (one with a rag-forge.config.ts at the root). The MCP server reads that config to know which vector store, embedding model, and judge to use.

Documentation

• Full docs: github.com/hallengray/rag-forge
• MCP overview: github.com/hallengray/rag-forge/blob/main/apps/docs/content/mcp/overview.mdx
• Issues: github.com/hallengray/rag-forge/issues

License

MIT — Femi Adedayo