Chonks

Local code retrieval for codebases too big to grep

Chonks indexes a large codebase into one local SQLite file and serves it to Claude Code, or any MCP harness, as retrieval tools. There is no LLM inside the pipeline. Retrieval is embeddings, BM25, and graph traversal, and the coding agent does the thinking.

The index is built, not live. Re-indexing is explicit and incremental, updating from the changed and deleted chunks, with a full rebuild above 20% churn. Every MCP response carries the index age, so an agent can see how old its answer is. There is no file watcher

Agentic search with grep, glob, and repo maps works well until the codebase outgrows it. A large codebase has tens of thousands of files, coupling that crosses language boundaries, and names that cannot be guessed, so at that scale an agent spends its context window walking directories. Chonks gives it sharper questions to ask:

• "Where is the thing that does X?" Hybrid semantic and keyword search over AST-boundary chunks (codebase_search).
• "Collect everything related to this." Iterative graph-expansion retrieval for multi-hop, cross-subsystem questions (codebase_research).
• "Who uses this symbol?" and "Where is it defined?" Structural lookups over a precomputed typed cross-reference graph (find_usages, `fi