rlm, created by SaschaOnTour, is an MCP server and development tool that treats a codebase as a queryable database to reduce ‘context rot’. The tool lets AI agents progressively narrow from project overviews to specific function bodies, perform AST-aware symbol lookups, and apply targeted edits with syntax validation. It is aimed at developers and AI-assisted agents working on large repositories who need fewer tokens and safer, more focused code modifications.

What tasks can you actually use it for?

rlm implements a query-driven workflow that replaces full-file reads with focused lookups, so agents can locate symbols, usages, and impact across a repository. Key behaviors include progressive disclosure to zoom from high-level structure to individual functions, AST-aware retrieval for precise symbol resolution, and surgical editing that swaps AST nodes instead of overwriting files. These capabilities map directly to tasks: code navigation, targeted refactors, and automated patch generation.

How accurate and safe are the edits it produces?

The tool applies a Syntax Guard that validates proposed changes against the repository’s Abstract Syntax Tree before writing to disk, which reduces the chance of introducing syntax errors. Because edits replace specific AST nodes, the approach narrows the edit surface and lowers the risk of unintended side effects. This safety model supports using AI agents for programmatic edits while keeping build-breaking syntax checks in place.

What inputs and repository sizes does it handle?

rlm keeps most of a codebase outside the active language model context and exposes only requested parts, which the developer reports cuts token usage by up to 90 percent. The tool uses a Rust-built semantic index for high-speed search and handles repositories that exceed standard context windows. It runs as an MCP-compatible server and accepts queries from MCP clients rather than ingesting entire files into the model context.

Does it fit into existing AI coding workflows?

rlm integrates with MCP clients such as Claude Desktop and Cursor, so it can slot into agent-driven IDE workflows that support MCP. The implementation in Rust targets memory safety and indexing performance, which helps when indexing large projects. The project is currently in Beta, so teams should plan for active development and evolving integration behavior while evaluating it alongside their existing agent toolchain.

A practical choice for AI-assisted development on large repositories

rlm is a practical option for developers and AI agents that need controlled, queryable access to large codebases; its local, source-available design removes external tracking concerns. Expect active changes while the project is in Beta, and treat generated edits as machine-assisted suggestions that benefit from human review before wide deployment.