LynxMCP

What lorenzo-cambiaghi/LynxMCP MCP server does

lorenzo-cambiaghi/LynxMCP MCP server gives MCP-compatible AI agents a local retrieval layer for software projects and related reference material. It can index source code, public documentation sites, and PDF files together, allowing searches to span multiple sources or target one source explicitly.

Code is divided according to syntax rather than arbitrary text windows. Tree-sitter supports more than 18 languages, including TSX, and lets the index represent complete functions, classes, and other code units. Search results can include the symbol, file and line location, score, and retrieved code. An outline mode returns signatures without full bodies for lower-cost candidate triage.

How it works

The retrieval pipeline combines dense embeddings with code-aware BM25 search and merges their results using reciprocal rank fusion. An optional cross-encoder reranker can further refine results. The first local build downloads an embedding model of about 130 MB.

When enabled, the knowledge graph records relationships such as callers, callees, subclasses, superclasses, imports, and neighboring symbols. These relationships support impact analysis, shortest paths, module summaries, and symbol context. The graph can also be exported as a single offline view for sharing or archiving.

A file watcher updates the index after saved changes, with updates described as taking roughly two seconds. Search and graph data are also available as rows through a local HTTP API, which can be joined with other local data in DuckDB or Coral. Documentation pages are fetched on demand; JavaScript-rendered single-page applications can be handled through optional headless Chromium support.

Setup and configuration

Install the command-line package with pipx or uv:

pipx install lynx-mcp
# or
uv tool install lynx-mcp

Create a project configuration with lynx manager init, then build the index with lynx build. The MCP client launches the service through the lynx serve --config command and a configuration file path. lynx manager ui provides guided setup, diagnostics, a query playground, and generated client configuration snippets.

Tools and capabilities

lorenzo-cambiaghi/LynxMCP MCP server exposes a fixed tool set, including:

• Hybrid and multi-query search, with optional source selection and outline results.
• Definition, usage, test, similar-code, symbol-description, and repository-overview queries.
• Graph operations for callers, callees, inheritance, imports, paths, neighbors, and status.
• Transitive impact analysis with affected tests and hop distances.
• File-level module summaries and offline graph exports.
• Searches limited to files changed against a base branch.
• Source listing, index status, source updates, and locally stored feedback reports.

Limitations and notes

The service is designed for local operation and the supplied material states that indexed data does not leave the machine. It requires an initial index build and local storage for the sources and embedding model. Knowledge-graph features are optional, so graph-dependent queries may not be available unless that feature is enabled. Documentation indexing may require optional browser support for JavaScript-rendered sites.

The project targets Python 3.10 or newer and is licensed under Apache 2.0. The README names Claude Code, Cursor, Windsurf, Antigravity, and generic stdio clients as supported client contexts; client-specific configuration details may vary.