LynxMCP: The Privacy-First "Long-Term Memory" Server for Cursor (AST-Aware + Graph + UI)

Showcase Built for Cursor
, , ,
1

Hey everyone! :waving_hand:

I’m excited to share LynxMCP, an open-source, 100% local MCP server I’ve been building to give Cursor an incredibly deep, structural understanding of massive codebases and external documentation.

Cursor’s native codebase indexing is great, but when working on huge enterprise projects, highly polymorphic code, or trying to use the absolute latest undocumented framework APIs, AI can still hallucinate or lose context. I built LynxMCP to solve exactly that.

GitHub Repo: https://github.com/lorenzo-cambiaghi/LynxMCP

:thinking: What does it do?

Lynx acts as a dedicated “brain” running locally on your machine. You point it at your local repositories, downloaded PDFs, or online framework documentations, and it exposes advanced semantic search and graph tools to Cursor via the Model Context Protocol.

:fire: Why use it alongside Cursor?

  • AST-Aware Chunking (Tree-sitter): Unlike naive text chunking, Lynx uses Tree-sitter (13+ languages supported). It understands where functions and classes begin and end, meaning Cursor gets intact, logical blocks of code as context instead of arbitrary cut-offs.

  • Knowledge Graph Layer: If you enable the graph mode, Lynx maps out your architecture. Cursor can literally ask: “What concrete classes implement this interface?” or “Show me the blast radius if I change this method.”

  • Unified Knowledge Base (Code + Docs): You can add external Web Documentation (like React, Unity, etc.) and local PDFs (like hardware manuals or RFCs) as sources. Cursor will transparently search the absolute latest docs before writing code, completely eliminating API hallucinations.

  • 100% Local & Privacy-First: Embeddings are computed on your CPU using an open model, and ChromaDB stores vectors on your local disk. Not a single byte of your proprietary code leaves your machine.

:control_knobs: The best part: LynxManager (Web UI)

You don’t need to fight with JSON configs or CLI commands. Lynx comes with a beautiful local web dashboard (LynxManager).

You don’t even need to touch a terminal or install Python to try it out. I’ve built thin native bootstrapper apps for macOS and Windows. You just double-click the installer, and a web panel opens up. From there, you can add your codebase paths, scrape web docs, test semantic queries in the Playground, and literally copy-paste the generated Cursor JSON configuration with one click.

:hammer_and_wrench: How to try it

Check out the GitHub link above! On the main page, under the “Install Lynx” section (or directly in the Releases tab), you’ll find the .dmg and .exe standalone installers. Just download, double-click, and let the UI guide you.

I’d love to get feedback from power users in this community. Let me know what you think, what languages you’d like to see better supported, or any crazy use-cases you find for the Graph layer!

Happy coding! :high_voltage:

3

Hey, looks great, thanks for sharing this with the community. AST-aware chunking via Tree-sitter and a graph layer for requests like “blast radius when changing a method” is a really interesting approach, especially for large monorepos. The fact that everything runs locally and embeddings are computed on CPU is also a big plus for users working with proprietary code.

A couple thoughts, if you’re interested:

  • It’s awesome that there’s a Playground to test semantic requests before hooking it up to Cursor, that really lowers the barrier to entry.
  • Let us know how indexing behaves on truly large codebases later on, like initial run time and ChromaDB size on disk. The community usually loves those numbers.

I’ll leave the thread open so others can test and share feedback. If you ship updates or add new languages in Tree-sitter, feel free to post here.