CodeRLM applies the Recursive Language Model (RLM) pattern to codebases. A Rust server indexes a project's files and symbols via tree-sitter, then exposes a JSON API that LLM agents query for targeted context — structure, symbols, source code, callers, tests, and grep. Instead of loading an entire codebase into context or relying on heuristic file scanning, the agent asks the server for exactly what it needs.

An integrated Claude Code skill (plugin/skills/coderlm/) wraps the API with a Python CLI and a structured workflow, so Claude Code can explore unfamiliar codebases without reading everything into context.

The RLM pattern treats a codebase as external data that a root language model can recursively examine and decompose:

1. Index — The server walks the project directory (respecting .gitignore), parses every supported file with tree-sitter, and builds a symbol table with cross-references.
2. Query — The agent queries the index: search symbols by name, list functions in a file, find callers of a function, grep for patterns, retrieve exact source code.
3. Read — The server returns the exact code requested — full function implementations, variable lists, line ranges — so the agent never guesses.

This replaces the typical glob/grep/read cycle with precise, index-backed lookups.

This project builds on two prior works:

• "Recursive Language Models" by Alex L. Zhang, Tim Kraska, and Omar Khattab (MIT CSAIL, 2025). The paper introduces the RLM framework for processing inputs far beyond model context windows by treating extended prompts as external data that the model recursively examines.

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

• brainqub3/claude_code_RLM — A minimal RLM implementation for Claude Code by brainqub3 that applies the pattern to document processing via a persistent Python REPL. CodeRLM adapts this approach from documents to codebases, replacing the Python REPL with a purpose-built Rust server and tree-sitter indexing.

server/                          Rust server (the only built artifact)
plugin/                          Self-contained Claude Code plugin
  plugin/skills/coderlm/         Skill definition + Python CLI wrapper
  plugin/hooks/                  Claude Code hooks (SessionStart, UserPromptSubmit, PreCompact, Stop)
  plugin/commands/               Slash command definitions
  plugin/scripts/                Hook scripts (session lifecycle)
  plugin/.claude-plugin/         Plugin manifest (plugin.json)
.claude-plugin/                  Marketplace manifest (points to plugin/)

The fastest way to get started. Requires Rust toolchain for building the server.

# 1. Install the plugin (registers skill, hooks, and slash commands)
claude plugin install github:JaredStewart/coderlm

# 2. Build the server
cd ~/.claude/plugins/cache/coderlm/coderlm/latest/server
cargo build --release

# 3. Start the server (in a separate terminal)
cargo run --release -- serve

# 4. Restart Claude Code — the SessionStart hook will auto-initialize

After installation, the /coderlm skill is available in every Claude Code session. The UserPromptSubmit hook guides Claude to use it automatically for code exploration tasks.

# 1. Clone the repository
git clone https://github.com/JaredStewart/coderlm.git
cd coderlm

# 2. Build the server
cd server && cargo build --release

# 3. Start the server
cargo run --release -- serve /path/to/your/project

# 4. (Optional) Run as a daemon
./server/coderlm-daemon.sh start
./server/coderlm-daemon.sh status
./server/coderlm-daemon.sh stop

curl http://127.0.0.1:3000/api/v1/health
# → {"status":"ok","projects":0,"active_sessions":0,"max_projects":5}

Once the server is running, invoke the skill:

/coderlm query="how does authentication work?"

Or use the CLI directly:

python3 plugin/skills/coderlm/scripts/coderlm_cli.py init
python3 plugin/skills/coderlm/scripts/coderlm_cli.py search "handler"
python3 plugin/skills/coderlm/scripts/coderlm_cli.py impl run_server --file src/main.rs

claude plugin update coderlm
# Rebuild the server after updating
cd ~/.claude/plugins/cache/coderlm/coderlm/latest/server
cargo build --release

When installed, CodeRLM gives Claude Code:

• /coderlm skill — Structured workflow for codebase exploration (init → structure → search → impl → callers → synthesize)
• SessionStart hook — Auto-detects a running server and initializes sessions
• UserPromptSubmit hook — Guides Claude to use indexed lookups instead of glob/grep/read
• Zero Python dependencies — The CLI wrapper uses only the Python standard library

coderlm-server serve [PATH] [OPTIONS]

Options:
  -p, --port <PORT>              Port to listen on [default: 3000]
  -b, --bind <ADDR>              Bind address [default: 127.0.0.1]
      --max-file-size <BYTES>    Max file size to index [default: 1048576]
      --max-projects <N>         Max concurrent indexed projects [default: 5]

All file types appear in the file tree and are searchable via peek/grep, but only the above produce parsed symbols.

All endpoints under /api/v1/. See server/REPL_to_API.md for the full endpoint reference with curl examples.

MIT