CodeRLM

CodeRLM is a Rust server that indexes a codebase and exposes it to LLM agents as a queryable JSON API. It parses supported files with tree-sitter, builds a symbol table with cross-references, and answers targeted questions - project structure, symbol search, function implementations, callers, tests, and grep - so an agent requests exactly the code it needs instead of loading a repository into context or scanning files heuristically.

The approach applies the Recursive Language Model pattern, published by Alex L. Zhang, Tim Kraska, and Omar Khattab, which treats a large input as external data that a root model recursively examines. CodeRLM adapts an earlier document-oriented implementation from documents to codebases, replacing a Python REPL with the indexing server. Indexing respects .gitignore, and the server caps file size and the number of concurrently indexed projects.

A companion plugin wraps the API for Claude Code: a /coderlm skill with a structured exploration workflow (init, structure, search, implementation, callers, synthesize), session lifecycle hooks that auto-initialize and steer the agent toward indexed lookups, and a Python command-line wrapper that uses only the standard library. A generator produces equivalent configuration for other platforms, and the CLI can be used directly against a running server.