grafyx-mcp

Grafyx

Real-time codebase understanding for AI coding assistants.

What is Grafyx?

AI coding tools read raw files with zero architectural understanding -- they don't know what calls what, which classes inherit from where, or how your modules connect. Grafyx fixes this by parsing your entire codebase into a full relationship graph using Graph-sitter (built on tree-sitter), then exposing that graph to any AI assistant through the Model Context Protocol (MCP). Your assistant can trace call chains, map dependencies, find related code by description, detect conventions, and understand your project's architecture -- all in real time, with a file watcher that keeps the graph current as you edit.

Quick Start

Claude Code

# Zero-install (recommended)
claude mcp add --scope user grafyx -- uvx --from grafyx-mcp grafyx

# Or install with pip first
pip install grafyx-mcp
claude mcp add --scope user grafyx -- grafyx

Cursor / Windsurf / Cline

Add to your MCP config file:

• Cursor: .cursor/mcp.json (project) or ~/.cursor/mcp.json (global)
• Windsurf: ~/.codeium/windsurf/mcp_config.json
• Cline: Cline MCP settings in VS Code

{
  "mcpServers": {
    "grafyx": {
      "command": "uvx",
      "args": ["--from", "grafyx-mcp", "grafyx"]
    }
  }
}

VS Code (GitHub Copilot)

Add to .vscode/mcp.json:

{
  "servers": {
    "grafyx": {
      "command": "uvx",
      "args": ["--from", "grafyx-mcp", "grafyx"]
    }
  }
}

Using pip instead of uvx? Replace the command with: "command": "grafyx" (no args needed).

Available Tools

Tool	Description
get_project_skeleton	Full project structure with stats per module
get_function_context	Everything about a function: callers, callees, deps
get_file_context	File contents, imports, dependencies
get_class_context	Class methods, inheritance, usages
find_related_code	Natural language search across the codebase
find_related_files	Find files relevant to a feature by matching symbols
get_dependency_graph	Impact analysis: what depends on what
get_conventions	Detected coding patterns and conventions
get_call_graph	Call chain tracing upstream and downstream
refresh_graph	Force re-parse of the codebase
get_module_context	Symbols in a directory/package (intermediate zoom)
get_subclasses	Inheritance tree for a base class
get_unused_symbols	Dead code detection
set_project	Switch the served project at runtime

How It Works

Your AI Assistant
       |
       | MCP Protocol (stdio)
       v
  +-----------+
  |  Grafyx   |  FastMCP server with 14 tools
  |  Server   |
  +-----------+
       |
  +-----------+     +-----------+     +-------------+
  |  Graph    |---->|  Search   |     | Convention  |
  |  Engine   |---->|  Engine   |     | Detector    |
  +-----------+     +-----------+     +-------------+
       |
       v
  +-----------+
  |  Graph-   |  Tree-sitter based parsing
  |  sitter   |
  +-----------+
       |
  +-----------+
  |  Watchdog |  File watcher for live updates
  +-----------+

• Startup -- Grafyx detects languages in your project and parses all source files into a semantic graph via Graph-sitter.
• Serving -- The FastMCP server exposes 14 tools over stdio. Your AI assistant calls them as needed.
• Live updates -- Watchdog monitors file changes. When you save, the graph is automatically re-parsed after a short debounce.

ML-augmented search

Grafyx's find_related_code uses a pretrained code embedding model (default: jinaai/jina-embeddings-v2-base-code, Apache-2.0, 161M params) running on CPU via ONNX through fastembed. The model is downloaded on first use and cached locally — no GPU, no daemon, no cloud calls.

Since 0.2.1, fastembed is a hard dependency, so the default install already includes the encoder — no extra needed.

Benchmark (0.2.0, 278 docstring→function queries across FastAPI + Django):

Encoder	nDCG@10	MRR@10	p50 latency
jina-v2 (default)	0.787	0.741	~1.5 s
coderankembed	0.663	0.623	~1.3 s
tokens-only (no fastembed)	0.335	0.297	~0.9 s

The default encoder more than doubles retrieval quality over plain source-token search (+135% nDCG@10).

Full breakdown + per-query JSONL: docs/benchmarks/0.2.0/.

Switch encoders via the GRAFYX_ENCODER env var:

• jina-v2 (default) — Apache-2.0, fastembed-native, ~150 MB. Wins on accuracy; recommended unless you have a specific reason to switch.
• coderankembed — MIT, 137M, ONNX-int8, ~140 MB. Lower latency but ~12 nDCG@10 points behind jina-v2 in our eval. Hosted at Bilal7Dev/grafyx-coderankembed-onnx.

Supporting numpy-only MLPs (~5 MB total weights, bundled in the wheel):

• M1 Relevance ranker — 33-feature MLP that re-ranks the encoder's top candidates using structural signals (caller count, name overlap, exports).
• M3 Source token filter — suppresses noise tokens (imports, strings, magic methods) from full-text search.
• M4 Symbol importance — weights symbols by caller count, exports, and structural signals.
• Gibberish detector — character-bigram MLP that blocks nonsense queries before they hit the index.

Reproducible benchmarks against FastAPI, Django, and Home Assistant ship in benchmarks/ (python -m scripts.run_all).

Supported Languages

Language	Extensions
Python	.py, .pyi
TypeScript	.ts, .tsx
JavaScript	.js, .jsx

Languages are auto-detected. To specify manually:

grafyx --languages python,typescript

Options

grafyx [OPTIONS]

  --project PATH       Project to analyze (default: current directory)
  --languages LANGS    Comma-separated languages (default: auto-detect)
  --ignore PATTERNS    Additional directories to ignore
  --no-watch           Disable file watching
  --verbose, -v        Debug logging
  --version            Show version

Default ignored: node_modules, .git, __pycache__, .venv, venv, .env, dist, build, .tox, .mypy_cache, .pytest_cache, .ruff_cache, egg-info, .eggs, .next, .nuxt, coverage, .coverage, .nyc_output

Multi-Agent Support

Grafyx works with agent teams. A single Grafyx instance serves all agents connected to the same project. When one agent modifies code, the file watcher updates the graph automatically, so other agents immediately see the changes.

Contributing

git clone https://github.com/bilal07karadeniz/Grafyx.git
cd Grafyx
pip install -e ".[dev]"
pytest

Troubleshooting

Windows: Graph-sitter requires Linux. Use WSL and configure your MCP client to launch via wsl:

{
  "mcpServers": {
    "grafyx": {
      "command": "wsl",
      "args": ["-e", "bash", "-c", "source ~/your-venv/bin/activate && grafyx"]
    }
  }
}

License

MIT -- see LICENSE for details.