abivan-tech/zvec-mcp

This project is a local server that scans the source code files in a project, builds a searchable index of them, and exposes that index to AI coding tools through a standard protocol called MCP (Model Context Protocol). The practical result is that an AI assistant can run a natural-language search against your actual codebase and get back relevant code snippets, rather than relying only on what it can see in open editor tabs. When you start the bridge, it walks through your project directory and reads all the recognized source files, JavaScript, TypeScript, Python, Go, Java, Kotlin, Rust, and several others. It splits the text into chunks and converts each chunk into a numerical embedding using a small language model from Hugging Face that runs locally on your machine. Those embeddings go into a local database file stored at a path called .zvec/knowledge.db inside your project. After the initial scan, the bridge watches for file changes and updates the index automatically as you work. The bridge exposes four tools that an MCP-compatible AI client can call: one to search the index with a natural-language query, one to initialize or rebuild the full index, one to index a single file immediately, and one to check the current database status. When a search query comes in, the bridge generates an embedding for the query, runs a vector search to find the closest matches, and then applies additional keyword and path filtering before returning results. Installation requires Node.js 18 or newer and a standard npm install. The bridge downloads a small local model on first use, so an internet connection is needed that one time. After that it works entirely offline. Configuration is a short JSON snippet that points an MCP client at the bridge script and tells it which project directory to index. The README is sparse on advanced configuration, but the core use case is clear: any MCP-compatible AI tool can use this bridge to search your codebase semantically rather than by exact keyword matching.