Claude Code’s missing code-search layer, written in Rust

What it does

CodeSeek is a Rust CLI that parses your codebase with Tree-sitter, extracts functions and call relationships, and builds a hybrid search index mixing dense vector embeddings, BM25 keyword search, and serialized call graphs. It then exposes these indexes as native MCP tools—codeseek_search, codeseek_callers, and codeseek_callees—directly inside Claude Code and Codex CLI. Every command is a standalone process; there is no background daemon.

The interesting bit

The search pipeline is unusually elaborate for a local tool: LanceDB handles vector similarity, Tantivy handles BM25, results are fused with reciprocal rank fusion, and a Qwen3 cross-encoder reranker re-scores the top candidates—falling back to graph-based name search if the embedding API is unreachable. The project also ships via npm, but the package is just a thin TypeScript wrapper that downloads the correct Rust binary from GitHub Releases.

Key highlights

• Hybrid retrieval stack: dense vectors (LanceDB), sparse tokens (Tantivy BM25), RRF fusion, and a cross-encoder reranker, with call-graph fallback.
• Seven-language AST parsing (Rust, Python, JavaScript, TypeScript, Go, C/C++, Java) with incremental re-indexing via MD5 hashes.
• Native MCP integration: registers tools into ~/.claude.json and ~/.codex/config.toml so agents can query symbols and trace call chains.
• Standalone process model—no daemon, no HTTP server—stores indexes in ~/.codeseek/<project_hash>/.
• Optional git hooks trigger automatic re-index after commits and merges.

Caveats

• Semantic search and reranking require external API tokens (default configuration points to SiliconFlow); offline use is limited to graph-based symbol search.
• No Windows support is listed; binaries are provided only for macOS (arm64, x64) and Linux (x64).
• Building from source requires manually installing protoc first.

Verdict

Worth a look if you live in Claude Code or Codex CLI and want your agent to navigate large, polyglot repos with actual structure rather than filename hallucinations. Skip it if you need a fully offline semantic search stack or you are on Windows.