How to Configure the MCP Server for Claude Code Integration in Code-Graph-RAG

Configure the MCP server by setting the TARGET_REPO_PATH environment variable to your repository's absolute path, then register it with Claude Code using claude mcp add --transport stdio to enable semantic codebase queries.

The Model Context Protocol (MCP) server in the vitali87/code-graph-rag repository acts as the bridge between Claude Code and your codebase. When properly configured, this integration allows Claude Code to perform semantic searches, navigate AST relationships, and modify files through natural language commands. The configuration process centers on environment variables that define your target repository and LLM backend credentials.

Core Architecture and Configuration Files

The MCP server implementation follows a strict initialization sequence defined across three critical files. In codebase_rag/mcp/server.py, the entry point handles environment resolution, graph initialization, and request routing. The system first resolves the target repository path using a fallback chain defined at line 44:

repo_path = os.environ.get(cs.MCPEnvVar.TARGET_REPO_PATH) or settings.TARGET_REPO_PATH

Constant definitions reside in codebase_rag/constants/mcp.py, which contains the MCPEnvVar enumeration including the TARGET_REPO_PATH key. Default values and the Pydantic settings model live in codebase_rag/config.py, where TARGET_REPO_PATH defaults to "." (the current working directory). Tool descriptions that populate the MCP interface are defined in codebase_rag/tools/tool_descriptions.py and registered through the MCPToolsRegistry.

Required Environment Variables

The server reads four critical variables at startup to configure the target repository and LLM backend:

Variable	Source	Purpose	Example Value
TARGET_REPO_PATH	codebase_rag/constants/mcp.py	Absolute path to the codebase for analysis	/home/user/myproject
CYPHER_PROVIDER	Environment	LLM provider for graph queries	openai or google
CYPHER_MODEL	Environment	Specific model for embeddings/generation	gpt-5.6-luna
CYPHER_API_KEY	Environment	API authentication key	sk-...

If TARGET_REPO_PATH is omitted, the server falls back to the directory from which Claude Code was launched, as implemented in the fallback logic of codebase_rag/mcp/server.py.

Registering the Server with Claude Code

Installation via pip

For packaged installations, use the claude mcp add command with the stdio transport (required for Claude Code CLI integration):

claude mcp add --transport stdio code-graph-rag \
  --env TARGET_REPO_PATH=/absolute/path/to/your/project \
  --env CYPHER_PROVIDER=openai \
  --env CYPHER_MODEL=gpt-5.6-luna \
  --env CYPHER_API_KEY=your-api-key \
  -- code-graph-rag mcp-server

Running from Source

When executing from the repository directory, use uv run to manage dependencies:

claude mcp add --transport stdio code-graph-rag \
  --env TARGET_REPO_PATH=/absolute/path/to/your/project \
  --env CYPHER_PROVIDER=openai \
  --env CYPHER_MODEL=gpt-5.6-luna \
  --env CYPHER_API_KEY=your-api-key \
  -- uv run --directory /path/to/code-graph-rag code-graph-rag mcp-server

Dynamic Path Configuration

To analyze the current working directory dynamically, pass the environment variable at runtime:

cd /path/to/your/project
claude mcp add --transport stdio code-graph-rag \
  --env TARGET_REPO_PATH="$(pwd)" \
  --env CYPHER_PROVIDER=google \
  --env CYPHER_MODEL=gemini-3.5-flash-lite \
  --env CYPHER_API_KEY=your-google-api-key \
  -- uv run --directory /absolute/path/to/code-graph-rag code-graph-rag mcp-server

Available MCP Tools

Once connected, the server exposes tools defined in codebase_rag/tools/tool_descriptions.py through the MCPHandlerType routing system. Key capabilities include:

• semantic_search – Query the codebase using natural language to find relevant functions and classes
• list_projects – Enumerate available projects in the graph database
• write_file – Modify existing files or create new ones within the target repository

Requests are routed through the transport layer (stdio for Claude Code) and return serialized MCPResultType objects. For example, a semantic search request:

{
  "tool": "semantic_search",
  "params": {
    "question": "Find all functions that send an email"
  }
}

Summary

• Set TARGET_REPO_PATH to the absolute path of your target repository; the server reads this from codebase_rag/constants/mcp.py and falls back to codebase_rag/config.py defaults if unset
• Configure LLM variables (CYPHER_PROVIDER, CYPHER_MODEL, CYPHER_API_KEY) to enable graph queries and embeddings
• Use stdio transport when registering with Claude Code via claude mcp add --transport stdio for proper CLI integration
• Tool definitions are automatically loaded from codebase_rag/tools/tool_descriptions.py into the MCPToolsRegistry when the server initializes

Frequently Asked Questions

What happens if I don't set TARGET_REPO_PATH?

If the environment variable is omitted, the server executes the fallback logic in codebase_rag/mcp/server.py (line 44), defaulting to the current working directory from which Claude Code was launched. This default is defined in codebase_rag/config.py as TARGET_REPO_PATH = ".".

Can I switch from stdio to HTTP transport for Claude Code?

While the server implementation in codebase_rag/mcp/server.py supports both transports via the --transport argument, Claude Code specifically requires stdio for MCP integration. The HTTP transport option exists for other MCP clients, but stdio is mandatory for Claude Code CLI connections.

Where are the tool descriptions stored?

Human-readable tool descriptions reside in codebase_rag/tools/tool_descriptions.py. This file populates the MCPToolsRegistry with definitions for tools like semantic_search and write_file, which Claude Code displays in its interface.

Does the server handle Memgraph initialization automatically?

Yes. After resolving TARGET_REPO_PATH, the server automatically builds or incrementally updates the code graph in Memgraph. You do not need to manually populate the graph database; the server indexes the repository upon startup using the configured LLM provider for embeddings.

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