# NeuroMesh

A local-first MCP context engine that indexes a repository into a graph, routes a prompt to the relevant symbols, and sends the model a compact evidence packet instead of whole files.

NeuroMesh targets a common failure mode where an editor ships whole, thousand-line files to a model for a simple question, spending tokens on code that is never touched and burying the relevant part among unrelated helpers -- a problem it argues vector RAG and static code graphs only partly fix, since both still end up pasting full files into the prompt.

It keeps functions a query needs fully expressed while collapsing unused code into a one-line, reversible marker recording what was folded, so an agent still sees a file's signatures, imports and neighbors without paying to read every private helper, and can expand a specific fold back into memory with a dedicated tool when it is actually needed. It works as an MCP server with Cursor, VS Code, Claude, Codex, OpenCode, Antigravity, Kilo, Trae, Windsurf and Zed, and is written in Rust under the MIT license.

Website: https://pinoox.github.io/neuromesh/