What is an MCP server and why is it needed?

MCP (Model Context Protocol) is a protocol that enables AI agents to communicate with external tools. This MCP server enables agents to perform semantic search across a codebase. While built-in grep tools have a false negative rate reaching 50%, hybrid search improves accuracy by up to 30%.

Why was LanceDB chosen?

LanceDB is serverless and file-based, requiring no separate database server. Its FTS (full-text search) support enables hybrid search. Alternatives like PostgreSQL+pgvector or Qdrant require additional infrastructure.

What does Tree-sitter AST chunking mean?

Tree-sitter is a parser that converts source code into an abstract syntax tree (AST). AST-aware chunking splits code by functions and classes rather than by lines. This ensures each chunk forms a semantically meaningful unit, improving search accuracy.

Which AI agents does this server work with?

It works with any agent that supports the MCP protocol. Starting with Claude Code, other coding agents with MCP support can also use this server. Installation takes a single command.

Local Semantic Code Search: An MCP Server for AI Agents

A local semantic code search MCP server that solves AI coding agents' search problems using Tree-sitter AST chunking and hybrid search. Development process, architectural decisions, and real-world results.

Jan 20, 2026 Updated: Jun 4, 2026

TL;DR

AI coding agents make systematic errors in codebase search: 50% false negatives, 54-59% token waste, diff accuracy dropping to 60%. To solve these problems, I built a local MCP server using Tree-sitter AST chunking, LanceDB hybrid search, and sentence-transformers. Result: up to 30% accuracy improvement with hybrid search, 52/52 tests passing.

View Subscription Sign In