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/)

• Rust toolchain — required to build the server (rustup recommended)
• Python 3 — required for the CLI wrapper (stdlib only, no pip packages)

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

# Start the server (in a separate terminal or as a daemon)
cargo run --release -- serve

# Or run as a daemon
./coderlm-daemon.sh start
./coderlm-daemon.sh status
./coderlm-daemon.sh stop

Verify:

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

# Add the marketplace source first, then install the plugin
claude /plugin marketplace add JaredStewart/coderlm
claude plugin install coderlm

After installation, the /coderlm skill is available in every session. The SessionStart hook auto-initializes and the UserPromptSubmit hook guides Claude to use indexed lookups.

Install the generator:

uv tool install coderlm --from git+https://github.com/JaredStewart/coderlm.git

Generate for your platform:

coderlm --platform cursor
coderlm --list                    # see all supported platforms

Or run without installing:

uvx --from git+https://github.com/JaredStewart/coderlm.git coderlm --platform cursor

Or from a cloned repo:

python3 plugin/generate.py --platform cursor

Use --list to see all platforms, --dry-run to preview, --clean to remove generated files, and --platform all for everything.

Once the server is running, invoke the skill (Claude Code) or use the CLI directly:

# Claude Code
/coderlm query="how does authentication work?"

# Direct CLI usage
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 Code plugin
claude plugin update coderlm

# Other platforms — pull and regenerate
git pull
python3 plugin/generate.py --platform cursor

Rebuild the server after any update:

cd 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]

Languages with tree-sitter support produce full symbol tables (functions, classes, methods, callers, variables). SQL uses regex fallbacks for variable and definition detection. All file types appear in the file tree and are searchable via peek/grep.

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

MIT