# rag-mcp-server

A lightweight MCP server that exposes an existing Anthropic documentation RAG service as tools for Claude Desktop.

## What this is

This project wraps the `anthropic-docs-rag` API (a separately running RAG service) and makes it accessible to Claude Desktop via the Model Context Protocol (MCP). Claude Desktop can then call `search_anthropic_docs` or `list_available_topics` as native tools during a conversation — without any copy-pasting or manual API calls.

The actual intelligence (embedding, retrieval, answer generation) lives entirely in the RAG service. This server is purely a protocol adapter.

## Architecture

```
Claude Desktop
     │
     │  stdio (stdin/stdout)
     ▼
MCP Server  (src/server.py — FastMCP)
     │
     │  HTTP POST /ask
     ▼
RAG Service  (localhost:8002 — anthropic-docs-rag)
     │
     ├──▶ ChromaDB  (vector store)
     └──▶ Claude API  (answer generation)
```

**Transport: stdio vs. HTTP/SSE**

This server uses **stdio transport** — Claude Desktop launches the Python process directly and communicates over stdin/stdout. It's the standard for local MCP servers: simple, no port conflicts, no auth needed.

For remote or production MCP servers (shared across multiple users or machines), you'd switch to **HTTP/SSE transport**, where the MCP server runs as a persistent web service and clients connect via Server-Sent Events. FastMCP supports both; only the `mcp.run()` call changes.

## Tools

| Tool | Signature | Description |
|------|-----------|-------------|
| `search_anthropic_docs` | `(query: str) -> str` | Sends a question to the RAG service and returns the generated answer |
| `list_available_topics` | `() -> list[str]` | Returns the static list of topics covered in the indexed documentation |

## Prerequisites

The `anthropic-docs-rag` service must be running on port 8002 before starting this server or using the tools in Claude Desktop. This MCP server will start successfully either way, but tool calls will fail with a connection error if the RAG service is unavailable.

## Setup

```bash
source .venv/bin/activate
pip install -r requirements.txt

# Optional: override the RAG service URL
cp .env.example .env
```

## Claude Desktop Configuration

Add this to `~/Library/Application Support/Claude/claude_desktop_config.json`:

```json
{
  "mcpServers": {
    "anthropic-docs-rag": {
      "command": "/absolute/path/to/rag-mcp-server/.venv/bin/python",
      "args": ["/absolute/path/to/rag-mcp-server/src/server.py"],
      "env": {
        "RAG_SERVICE_URL": "http://localhost:8002"
      }
    }
  }
}
```

Restart Claude Desktop — the two tools will appear automatically in the tool list.

## Tests

```bash
pytest tests/ -v
```

3 tests, all HTTP calls to the RAG service are mocked with `unittest.mock`.

## Note

This is a learning project for understanding MCP — how tools are defined, how Claude Desktop discovers and calls them, and how stdio transport works. It is not a standalone production system. All the interesting logic (RAG pipeline, embeddings, vector search) lives in the `anthropic-docs-rag` service.