CodeRLM · open source · MIT

AI Writes Code.
It Doesn't Understand Your System.

CodeRLM is an open-source code index that gives AI coding agents precise, on-demand access to symbols, callers, and implementations — instead of globbing directories and guessing.

Install CodeRLM → View on GitHub

$ claude /plugin marketplace add JaredStewart/coderlm

The problem

Agents read code they don't need. Then miss the code they do.

A typical coding agent opens dozens of files to answer one question — and still doesn't know that authenticate_user() is called from three places it never opened. CodeRLM skips the grep-guess-read loop entirely.

Without CodeRLM

glob, grep, read, repeat

01globsrc/**/*.py

02readdozens of files into context

03grep"authenticate" — many matches, no structure

04readmore files, still looking

05guesswhich match is the right one

06retrywith a different grep

approachread everything, hope

signal-to-noisemostly noise

With CodeRLM

one call, exact answer

01search"authenticate" — typed symbols, with locations

02callersof authenticate_user — the exact files

03implof authenticate_user — the source, not a guess

—no glob. no scan. no guess.

approachask for what you need

signal-to-noiseorder-of-magnitude better

What CodeRLM provides

A real-time symbol index, served as a JSON API.

Symbol table with cross-references

Every function, class, and method indexed with its callers and implementations. Search by name, jump straight to the source — no globbing required.

Tree-sitter parsing

Project files parsed into a structured symbol table. Edits re-index incrementally, so the index reflects what's actually on disk.

Agent-native CLI & API

search, callers, impl, structure, grep. Composable primitives a model can chain — built for agents, not for humans clicking in an IDE.

Claude Code plugin included

A bundled plugin with SessionStart, UserPromptSubmit, and PreCompact hooks that guide Claude to use indexed lookups instead of file scans.

Works with other tools

A generator emits the right config files for Cursor, Windsurf, Copilot, Gemini, and Codex. Pick your editor; CodeRLM travels with you.

Local-first, self-hosted

The server runs on your machine. The index doesn't leave it. MIT-licensed, single Rust binary, no telemetry.

agent ⟶ CodeRLM · localhost:3000

$ coderlm init .
{
  "session_id":    "s_4f7a2c",
  "project":       "platform-api",
  "files_indexed": 487,
  "symbols":       3214
}

$ coderlm search "authenticate"
[
  { "name": "authenticate_user", "kind": "fn",    "file": "src/auth.py:42" },
  { "name": "AuthMiddleware",    "kind": "class", "file": "src/middleware.py:18" },
  { "name": "verify_token",      "kind": "fn",    "file": "src/auth.py:91" }
]

$ coderlm callers authenticate_user --file src/auth.py
[
  "src/api/login.py:23",
  "src/api/refresh.py:14",
  "src/middleware.py:22"
]

$ # the agent now has exact locations. it reads three files, not forty.

Benchmarked support today

python

rust

javascript

In active development: typescript · go · java · scala

Origins

Built on the Recursive Language Model pattern.

CodeRLM applies the Recursive Language Model pattern to codebases. Instead of stuffing source files into a context window, a root model recursively examines an external index — asking only for the symbols, callers, and implementations it actually needs.

Zhang, A. L., Kraska, T., & Khattab, O. (2025). Recursive Language Models. arXiv:2512.24601.

The codebase adaptation builds on brainqub3/claude_code_RLM, a minimal RLM implementation for Claude Code that applies the pattern to documents via a Python REPL. CodeRLM swaps that REPL for a purpose-built Rust server with tree-sitter indexing — moving from documents to code.

Reinforcement Studio · early preview

Specs that live next to the code.

A managed workspace built on CodeRLM. Teams link their repos — schema, API, frontend, IaC — and author product specs alongside them. When the code drifts from the spec, you'll know.

Spec editor with code symbol links and diagrams
Bidirectional symbol ↔ spec references — drift detection
Agents that read the whole system, not just one repo
Built for teams of ≤ 50 — not enterprise procurement cycles

Email me about Studio Request a demo

Studio is exploratory and not yet generally available. The mockup is illustrative.

Specs

Auth · session model

Billing · pricing

Agent pipeline

Repos

platform-api

customer-web

db-schema

● in review

Auth: refresh-token rotation

Move refresh tokens from sticky-session storage to per-device records, rotating on every refresh.

authenticate_user()
  → src/auth.py:42

3 threads

Reviewer · 1hWhat about device fingerprinting?

Reviewer · 2hHow do we handle expired refresh?

✦ CodeRLM · 3hDrift: verify_token() now takes a device_id.

Get started

Give your agents the same index your IDE has.

CodeRLM is open source and self-hostable. Pick the path that matches your setup.

Claude Code plugin

The fastest path. Adds the /coderlm skill and the session hooks that guide Claude to indexed lookups.

$ claude /plugin marketplace add JaredStewart/coderlm
$ claude plugin install coderlm

Other AI tools

Cursor, Windsurf, Copilot, Gemini, Codex. The generator emits the right config for each.

$ uv tool install coderlm \
    --from git+https://github.com/JaredStewart/coderlm.git
$ coderlm --platform cursor

See --list for all supported platforms.

Build the server from source

Required for both paths above. Single Rust binary; no daemon dependencies.

$ git clone https://github.com/JaredStewart/coderlm.git
$ cd coderlm/server
$ cargo build --release
$ cargo run --release -- serve

Health check: curl localhost:3000/api/v1/health

Run as a daemon

Once built, manage the server in the background.

$ ./coderlm-daemon.sh start
$ ./coderlm-daemon.sh status
$ ./coderlm-daemon.sh stop

AI Writes Code.It Doesn't Understand Your System.