How to Start the Code‑Graph‑RAG MCP Server: A Complete Setup Guide

Launch the Code‑Graph‑RAG MCP server using the cgr mcp-server CLI command, which dispatches to either STDIO or HTTP transport handlers defined in codebase_rag/mcp.py, after starting a Memgraph instance with cgr daemon up.

The Model Context Protocol (MCP) enables AI assistants like Claude Code to query and modify codebases through natural language. The Code‑Graph‑RAG project (vitali87/code-graph-rag) exposes its graph‑powered RAG capabilities as an MCP server, allowing clients to index repositories and execute Cypher queries against a Memgraph database. To start the code-graph-rag MCP server, you must configure environment variables, launch the graph database, and invoke the correct transport mode.

Prerequisites and Environment Setup

Before starting the MCP server, ensure Memgraph is running and your environment is configured.

Memgraph Database: The MCP server requires an active Memgraph instance to store and query code graphs. Start the container using:

cgr daemon up

Required Environment Variables: The server expects several environment variables to be set, particularly TARGET_REPO_PATH pointing to the repository you want to analyze. When connecting via Claude Code, you must also provide LLM configuration for Cypher generation:

TARGET_REPO_PATH=/absolute/path/to/project
CYPHER_PROVIDER=openai
CYPHER_MODEL=gpt-4o  # or your preferred model

CYPHER_API_KEY=YOUR_API_KEY

Installation Options

Install the package using a Python package manager. The [treesitter-full,semantic] extras are required for full language parsing capabilities.

# Using uv (recommended)

uv tool install "code-graph-rag[treesitter-full,semantic]"

# Using pipx

pipx install "code-graph-rag[treesitter-full,semantic]"

Starting the MCP Server

The entry point for the server is implemented in codebase_rag/cli.py in the mcp_server function (around line 950). This function parses CLI arguments and dispatches to the appropriate transport handler in codebase_rag/mcp.py.

STDIO Transport (Default)

The STDIO transport is the default and simplest method, suitable for local MCP clients like Claude Code. It calls serve_stdio() from codebase_rag/mcp.py:

cgr mcp-server

HTTP Transport

For remote clients or web-based integrations, use the HTTP transport by calling serve_http(host, port):

cgr mcp-server --transport http --host 0.0.0.0 --port 8000

The dispatch logic in cli.py (lines 558‑566) routes to serve_stdio() when no transport flag is provided, otherwise invoking serve_http() with your specified host and port parameters.

Connecting Claude Code to the Server

Register the server with Claude Code using the claude mcp add command. This configuration uses STDIO transport and passes the required environment variables:

claude mcp add --transport stdio code-graph-rag \
  --env TARGET_REPO_PATH=/absolute/path/to/project \
  --env CYPHER_PROVIDER=openai \
  --env CYPHER_MODEL=gpt-4o \
  --env CYPHER_API_KEY=YOUR_API_KEY \
  -- cgr mcp-server

For the current working directory, substitute $(pwd) for the TARGET_REPO_PATH value.

Verifying the Server Connection

After registration, confirm the server is active by listing available tools:

claude mcp list

You should see tools including list_projects, index_repository, and query_code_graph, as defined in the MCP server implementation. These tools allow natural language queries against your codebase graph stored in Memgraph.

Summary

• Install Code‑Graph‑RAG with tree‑sitter and semantic extras via uv or pipx.
• Start Memgraph using cgr daemon up before launching the MCP server.
• Launch the server with cgr mcp-server for STDIO transport or add --transport http for remote access.
• Configure the TARGET_REPO_PATH environment variable to point to your target repository.
• Connect Claude Code using claude mcp add with the appropriate environment variables for Cypher generation.
• Verify functionality by running claude mcp list to see available graph querying tools.

Frequently Asked Questions

What transport protocols does the Code‑Graph‑RAG MCP server support?

The server supports STDIO (standard input/output) and HTTP transports. STDIO is the default and is used for local CLI clients like Claude Code, while HTTP is available for remote connections via the --transport http flag.

Why does the MCP server require Memgraph to be running?

The server queries and updates a graph database to provide code navigation and semantic search. Memgraph stores the abstract syntax tree (AST) and code relationships. Start it with cgr daemon up before launching the MCP server, or the initialization will fail.

Where is the MCP server logic implemented in the codebase?

The command‑line interface is defined in codebase_rag/cli.py in the mcp_server function, which dispatches to transport‑specific handlers. The actual server implementations—serve_stdio() and serve_http()—reside in codebase_rag/mcp.py.

Can I use a different LLM provider for Cypher query generation?

Yes. While the examples use OpenAI, you can configure any compatible provider by setting the CYPHER_PROVIDER, CYPHER_MODEL, and CYPHER_API_KEY environment variables when registering the server with your MCP client.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}