Codegraph

Big computer programs are made of tiny pieces that all talk to each other, like friends passing notes. Codegraph draws a map of all those friends and notes in Neo4j, calculates structural risk scores, finds similar functions with ML vector search, and exposes these intelligence capabilities to AI Assistants via MCP (Model Context Protocol).

1. Python AST Dependency Parser: Parses Python files, functions, docstrings, lines of code, call graphs, and file imports.
2. Impact Blast Radius Analysis: Predicts what functions or files will break if a target component is modified.
3. Risk Scoring Engine: Calculates composite risk scores based on caller count (in-degree), blast radius depth, lines of code, and Git commit churn.
4. ML Code Similarity Search: Uses TF-IDF and AST feature vector representations to find semantically and structurally similar functions.
5. MCP Server Integration: Exposes graph queries, impact analysis, risk scores, and similarity search as standardized tools for AI assistants like Claude Desktop, Cursor, or Antigravity.
6. Interactive Web Dashboard: FastAPI backend with Vis.js interactive dependency graph visualization, Risk Heatmaps, and ML Similarity Explorer.

docker compose up -d

• Browser: http://localhost:7474
• Credentials: neo4j / codegraph123

python -m venv venv
venv\Scripts\activate
pip install -r requirements.txt

python -m indexer index sample --reset

• Impact Analysis (What breaks if I change connect?):

  python -m indexer impact connect

• Risk Scoring (Which components are riskiest to touch?):

  python -m indexer risk --category functions

• ML Similarity Search (What functions are similar to login?):

  python -m indexer similar login

• Start Web Dashboard (Studio & Debugger):

  python -m indexer serve

  • Codegraph Studio UI: http://localhost:8000 (Interactive Graph Canvas, Code Inspector, Blast Radius Simulator, Risk Heatmap, ML Similarity Comparator)
  • Developer Debug Console: http://localhost:8000/debug (Direct Cypher query terminal & raw node inspection)

• Start MCP Server:

  python -m indexer mcp

Add Codegraph as an MCP server in your claude_desktop_config.json:

{
  "mcpServers": {
    "codegraph": {
      "command": "python",
      "args": ["-m", "indexer", "mcp"],
      "cwd": "C:/path/to/Codegraph"
    }
  }
}

Now you can ask your AI assistant questions like:

• "What breaks if I modify connect() in db.py?"
• "Find functions similar to login in my codebase."
• "Which parts of this codebase are highest risk to touch?"

Codegraph/
├── indexer/
│   ├── python_parser.py   # Python AST parser & docstring/snippet extractor
│   ├── loader.py          # Neo4j schema & batch loader with Git churn tracking
│   ├── impact.py          # Cypher impact & blast radius queries
│   ├── risk.py            # Risk scoring engine (centrality + churn + complexity)
│   ├── ml.py              # ML vector embedding & cosine similarity engine
│   ├── config.py          # Configuration defaults
│   └── __main__.py        # CLI interface
├── web/
│   ├── app.py             # FastAPI REST endpoints
│   └── static/            # Vis.js graph UI, Risk Heatmap, Similarity Explorer
├── mcp_server.py          # Model Context Protocol stdio server
├── tests/                 # Unit and integration test suite
├── docker-compose.yml     # Neo4j database definition
└── requirements.txt