// the find
zilliztech/claude-context
Code search MCP for Claude Code. Make entire codebase the context for any coding agent.
An MCP server that indexes your codebase into a vector database (Milvus/Zilliz Cloud) and exposes semantic + BM25 hybrid search to AI coding agents. It's from Zilliz, the company behind Milvus, so the vector database dependency is not accidental — this is partly a funnel for their managed cloud product. That said, self-hosted Milvus works too.
AST-based code chunking is a real differentiator over naive character splitting — it preserves function and class boundaries so retrieved chunks are actually usable. Merkle-tree-based incremental indexing means re-indexing after a small change doesn't re-embed the whole repo, which matters for large codebases. Hybrid BM25 + dense vector search gives better recall than either alone, particularly for identifier names that semantic search tends to miss. The core package is cleanly separated from the MCP layer, so you can use the indexing engine in your own tooling without pulling in the MCP scaffolding.
The hard dependency on an external vector database (Zilliz Cloud or a running Milvus instance) is a meaningful setup cost — there's no local-only path using something like SQLite-vec or Chroma, which would lower the barrier significantly. The embedding dependency defaults to OpenAI, so a fresh install needs two API keys before it does anything useful. The evaluation section claims ~40% token reduction but the methodology lives in a Python script and case studies on two GitHub issues, which is thin evidence for a headline claim. The Chrome extension is listed in the repo but marked as incomplete in the roadmap, so the directory tree is somewhat aspirational.