CodeGraph Smart MCP

Code Intelligence MCP Server for Claude — auto-indexes your codebase into a knowledge graph with GraphRAG + hybrid search, and opens a live dashboard.

Features

• Zero config — one command to index, search, and visualize any codebase
• GraphRAG — builds a knowledge graph of symbols, calls, imports, and clusters
• Hybrid search — BM25 + vector similarity (local embeddings, no API key needed)
• MCP Server — exposes tools to Claude Desktop for deep code understanding
• Live dashboard — Next.js + D3.js interactive graph visualization
• Multi-language — Python, JavaScript, TypeScript, JSX, TSX via Tree-sitter

Quick Start

npm install -g codegraph-smart-mcp

Then navigate to any project and run:

codegraph

That's it. First run indexes your code and opens the dashboard. Next runs are instant.

Requirements

• Node.js >= 18
• Python >= 3.9 (auto-detected, venv created automatically)

Usage

# Smart start — index + dashboard + open browser
codegraph /path/to/your/project

# Force re-index
codegraph --reindex

# Configure Claude Desktop (standby mode)
codegraph setup

# Index + configure Claude Desktop
codegraph init /path/to/project

# MCP stdio server (for Claude Desktop)
codegraph serve /path/to/project

# API server only (no browser)
codegraph start /path/to/project

# Check index status
codegraph status

# Diagnose installation
codegraph doctor

MCP Tools (for Claude)

| Tool | Purpose | |------|---------| | index_project | Index a directory on demand | | query | Hybrid BM25 + semantic search over symbols | | context | 360-degree view: callers, callees, file location | | impact | Blast radius — downstream/upstream traversal | | get_content | Read raw file content | | list_clusters | Functional module overview | | explore_process | Execution flow traces | | get_status | Check index and search status |

Claude Desktop Integration

After installing, run:

codegraph setup

This auto-configures Claude Desktop. Restart Claude Desktop, then ask:

"Index my project at /path/to/project"

Or pre-index a project:

codegraph init /path/to/project

Architecture

Your Project
    |
    v
Ingestion Pipeline (7 phases)
    |- Filesystem walk -> FileNodes
    |- Tree-sitter AST -> SymbolNodes + edges
    |- Cross-file import/call resolution
    |- Community detection (label propagation)
    |- Entry point detection + call chains
    |- Markdown/txt doc chunking
    |- Sentence-transformer embeddings
    |
    v
Storage: Neo4j (graph) + Qdrant (vectors)
    |
    v
Interfaces: MCP Server + REST API + Next.js Dashboard

Search Modes

• BM25 only — always available, no external dependencies
• Hybrid (BM25 + vectors) — requires Qdrant. Default local model: all-MiniLM-L6-v2 (~80MB, CPU)
• OpenAI embeddings — optional, set CODEGRAPH_OPENAI_API_KEY

Configuration

All settings use CODEGRAPH_ prefix:

CODEGRAPH_NEO4J_URI=bolt://localhost:7687
CODEGRAPH_NEO4J_PASSWORD=codegraph123
CODEGRAPH_QDRANT_MODE=disk            # memory | disk | server
CODEGRAPH_EMBEDDING_BACKEND=local     # local | openai
CODEGRAPH_OPENAI_API_KEY=sk-...       # required if backend=openai

License

MIT