vector-mcp-rust is a deterministic Model Context Protocol (MCP) server written in Rust. It gives AI coding agents hybrid code search, semantic AST-aware chunking, architecture decision record validation, and LSP-verified edits — with the entire index path running locally on CPU.
Why Rust
- Local AI privacy. Hugging Face
candle-coreruns quantized GGUF models (for example,Qwen2.5-0.5B-Instruct) on CPU, so indexing works fully offline and source code never leaves the machine. - Memory safety. The Rust
Droptrait reclaims C-allocated Tree-sitter AST memory deterministically — huge scans complete with zero garbage-collection pauses and zero leaks. - Hybrid search. LanceDB vector ANN (
IvfPq) and Tantivy BM25 run side by side, fused through Reciprocal Rank Fusion so lexical and semantic signals contribute to one ranked result.
Capabilities
- Semantic AST chunking for Go, Rust, TypeScript, JavaScript, PHP, and Python.
- The Fat Tool Pattern: exactly five super-tools, which prevents context exhaustion and hallucinated tool choice during agent sessions.
- A CDC watcher that validates ADRs against new code on every save.
- Cross-project multi-repo search across all indexed workspaces.
Safety model
Mutations are verified before they land. Every proposed patch is checked against a real language server — rust-analyzer for Rust, tsserver for TypeScript and JavaScript — and any patch that would produce compiler errors is rejected.
Rust · Hugging Face candle-core · Tree-sitter · LanceDB · Tantivy · rust-analyzer / tsserver