Start Here (Simple)¶

This page is the shortest path to understand RLM Code and start safely.

What RLM Code Is¶

RLM Code is a terminal app for running research experiments with language models.

It helps you:

• run recursive RLM workflows (/rlm run)
• run benchmark packs (/rlm bench ...)
• compare runs (/rlm bench compare ...)
• replay what happened (/rlm replay <run_id>)
• run coding-agent harness workflows (/harness run ...)

What RLM Code Is Not¶

RLM Code is not:

• a one-click product for non-technical users
• a guaranteed cheap tool (LLM calls can become expensive)
• a replacement for your own evaluation criteria
• fully safe if you force unsafe backend settings (exec)

What You Must Install¶

Required:

1. Python 3.11+
2. uv (recommended installer)
3. rlm-code package
4. At least one model route:
5. BYOK API key (OpenAI/Anthropic/Gemini), or
6. local model server (for example Ollama)

Recommended for safe execution:

1. Docker runtime (preferred default)
2. or Monty backend (pip install pydantic-monty) if you do not want Docker

Optional:

1. Apple container runtime (container CLI, macOS only, experimental)
2. cloud runtimes (Modal/E2B/Daytona) if needed

First Safe Session¶

uv tool install "rlm-code[tui,llm-all]"
rlm-code

In TUI:

/connect
/sandbox profile secure
/sandbox backend docker
/sandbox doctor
/rlm run "small test task" steps=4 timeout=30 budget=60
/rlm status

Use It as a Coding Agent (Simple)¶

You can use RLM Code like a coding assistant without running harness commands first.

Just connect once, then ask coding tasks directly in chat.

/connect

or

/connect acp

Then type normal prompts in chat, for example:

fix failing tests in this repo and explain the root cause

implement a parser for this config format and add unit tests

refactor this module for readability and keep behavior unchanged

Optional advanced mode:

• Use /harness run ... when you want explicit tool-loop control.
• Use /rlm run ... when you want explicit recursive experiment control.

Cost + Safety Warning¶

RLM experiments can trigger many model calls (especially recursive runs).

Always start with small limits:

• steps=4
• timeout=30
• budget=60
• small benchmark limits first (for example limit=1)

If a run is going out of control, stop it:

/rlm abort all

Or stop one run:

/rlm abort <run_id>

Use /rlm status to monitor the run and confirm whether it completed or was cancelled.

Fast Command Cheat Sheet¶