How to Set Up Claude Code with Code-Graph-RAG: Complete MCP Integration Guide

Code-Graph-RAG integrates with Claude Code through the Model Context Protocol (MCP), exposing a JSON-RPC endpoint that enables natural-language queries and AST-based code edits against a Memgraph-backed knowledge graph.

The vitali87/code-graph-rag repository provides a bridge between AI assistants and codebase intelligence through a graph-based retrieval system. Setting up Claude Code with Code-Graph-RAG allows you to query complex code relationships and execute refactoring operations using plain English instructions rather than manual navigation.

Understanding the MCP Architecture

The integration follows a three-tier architecture that decouples the AI client from the storage engine:

Claude Code (MCP client) ⇄ Code-Graph-RAG MCP server ⇄ Memgraph graph DB

When Claude Code sends a request, the Model Context Protocol (MCP) server in Code-Graph-RAG translates natural language into Cypher queries and executes them against the Memgraph database. This design is language-agnostic, meaning the same setup works for Python, TypeScript, Rust, Go, and other supported languages. The MCP layer is intentionally thin—it merely forwards JSON-RPC method calls to the underlying cgr commands that power the interactive REPL.

Installing Code-Graph-RAG and Dependencies

First, install the CLI tool with full Tree-sitter support to enable parsing across multiple languages. According to the repository documentation, the recommended method uses uv:

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

Alternatively, you can use pipx if you prefer:

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

This command installs the cgr executable along with semantic analysis capabilities and language parsers required for building the knowledge graph.

Initializing the Knowledge Graph

Before connecting Claude Code, you must index your target repository into Memgraph. Start the database daemon using Docker:

cgr daemon up

Then parse your codebase and populate the graph:

cgr start --repo-path /path/to/my/project --update-graph

This step analyzes the repository structure, builds an abstract syntax tree (AST) representation, and stores the relationships in Memgraph. The cgr start command uses the same underlying logic that the MCP server will later access when answering queries.

Launching the MCP Server

The MCP server implementation resides in cgr/__main__.py (or the entry point cgr/cli.py), which implements the mcp-server sub-command. You can start the server in two modes:

HTTP mode (default on port 8000):

cgr mcp-server

Stdio mode (for Claude Code's built-in client):

cgr mcp-server --stdio

The HTTP server exposes endpoints on localhost:8000 by default, while stdio mode communicates directly through standard input/output streams. Both modes provide the same JSON-RPC interface for querying the knowledge graph and submitting code edits.

Configuring Claude Code

With the server running, configure Claude Code to connect to the MCP endpoint. Detailed instructions are available in docs/claude-code-setup.md, but the process involves:

1. Opening Settings → MCP Server in Claude Code
2. Setting the Endpoint to http://localhost:8000 (or leaving blank for stdio)
3. Optionally specifying TARGET_REPO_PATH to default to a particular repository

Once connected, Claude Code can send natural-language prompts to the server, which generates Cypher queries, runs them against Memgraph via codebase_rag/workspaces/storage.py, and returns structured results.

Querying and Editing Code

After configuration, test the integration with graph-aware queries. For example:

"Show me all functions that call fetch_data"

The MCP server processes this by traversing the graph for call relationships and returns the matching function definitions. For code editing, you can request patches—the server builds an AST-based diff, applies it (optionally after confirmation), and updates the graph to reflect the changes.

Summary

• Code-Graph-RAG exposes functionality via the MCP protocol using a JSON-RPC/stdio interface that Claude Code consumes as a client.
• The installation requires uv tool install "code-graph-rag[treesitter-full,semantic]" to obtain the cgr CLI and language parsers.
• Infrastructure depends on Memgraph started via cgr daemon up, with repositories indexed using cgr start --repo-path --update-graph.
• The MCP server supports both HTTP (cgr mcp-server) and stdio (cgr mcp-server --stdio) transport modes, implemented in the CLI entry point.
• Configuration in Claude Code requires pointing to http://localhost:8000 or enabling stdio mode, as documented in docs/guide/mcp-server.md and docs/claude-code-setup.md.

Frequently Asked Questions

What is the Model Context Protocol (MCP) in Code-Graph-RAG?

The Model Context Protocol is a JSON-RPC interface that allows Claude Code to communicate with Code-Graph-RAG as an external tool. It standardizes how the AI client requests graph queries and code edits, enabling the server to handle Cypher generation and AST manipulation independently of the client implementation.

Can I use Code-Graph-RAG with editors other than Claude Code?

Yes, any MCP-compatible client can connect to the Code-Graph-RAG server. Because the protocol is standardized and language-agnostic, the same cgr mcp-server instance works with other AI assistants or IDE integrations that support MCP, provided they can send JSON-RPC requests over HTTP or stdio.

How does the MCP server handle code edits?

When you request a code change through Claude Code, the server constructs an AST-based diff of the proposed modification. It applies the patch to the filesystem (optionally requiring user confirmation) and then updates the Memgraph database to reflect the new code structure, ensuring the knowledge graph remains synchronized with the actual source files.

Where is the MCP server configuration documented?

The repository contains specific documentation in docs/guide/mcp-server.md for server command flags and transport options, and docs/claude-code-setup.md for client-side configuration. The high-level architecture overview appears in the README.md under the MCP Server section, explaining how the components interact to provide natural-language codebase interaction.

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