# rllm

> rllm is an open-source AI tool for ai models and llm tooling. Reinforcement-learning framework for training language agents on any harness. It is available on CLI and self-hosted.

- **Kind:** Tool
- **Pricing:** Open source
- **Platforms:** CLI, self-hosted
- **Website:** https://docs.rllm-project.com
- **Repository:** https://github.com/rllm-org/rllm
- **GitHub stars:** 5,921
- **License:** Apache-2.0
- **Categories:** AI models and LLM tooling, research
- **Listing updated:** 2026-11-14

## About

rLLM focuses on improving agents by training them, not just running them. A workflow engine runs many parallel agent instances to collect rollouts, a LiteLLM proxy routes requests and captures token IDs and logprobs, a transform pipeline groups trajectories, and a training backend (verl, tinker or fireworks) applies the policy update.

It can wrap CLI harnesses such as Claude Code, Codex and opencode or custom agents in LangGraph and the OpenAI Agents SDK, using sandboxes such as Docker, Daytona, Modal or local. Over 60 benchmarks are integrated, including SWE-bench and Terminal-Bench 2.0, and methods include GRPO, REINFORCE, RLOO and SFT.

The CLI covers model setup, eval, train and project scaffolding. It comes from the Berkeley Sky Computing Lab and is Apache-2.0.

## Features

- **Any harness:** Wrap CLI or custom agents with a decorator.
- **Switchable backends:** verl, tinker or fireworks with one flag.
- **60+ benchmarks:** rllm eval pulls and runs integrated benchmarks.
- **Several RL methods:** GRPO, REINFORCE, RLOO, SFT and distillation.
- **Sandbox support:** Docker, Daytona, Modal or local runs.
- **CLI workflow:** Setup, eval, train and init commands.

## Good for

- Training a coding agent with RL on SWE tasks
- Evaluating an agent across many benchmarks
- Switching between single-machine and multi-GPU training
- Reproducing published agent results

## FAQ

### Which training backends?

verl, tinker and fireworks.

### Does it need a particular agent framework?

No; it works with any framework with minimal code changes.

### Who builds it?

The Berkeley Sky Computing Lab.

### Is rllm free?

Yes. rllm is open source, so you can use and self-host it for free.

### What are the best alternatives to rllm?

The closest alternatives to rllm are GPT-3 Playground (OpenAI), ChatGPT (OpenAI), to teach, Heyday and ECC.

## Alternatives

- [GPT-3 Playground (OpenAI)](https://ai-powered.com/tools/gpt-3_playground_(openai)): GPT-3 Playground (OpenAI) is a free AI tool for copywriting.
- [ChatGPT (OpenAI)](https://ai-powered.com/tools/chatgpt_(openai)): ChatGPT (OpenAI) is a free AI tool for chat.
- [to teach](https://ai-powered.com/tools/to_teach): to teach is a freemium AI tool for productivity.
- [Heyday](https://ai-powered.com/tools/heyday): Heyday is an AI tool for productivity.
- [ECC](https://ai-powered.com/tools/ecc): ECC is a freemium AI tool for ai models and llm tooling.
- [Teach Anything](https://ai-powered.com/tools/teach_anything): Teach Anything is a free AI tool for research.
- [Perplexity AI](https://ai-powered.com/tools/perplexity_ai): Perplexity AI is a free AI tool for chat.
- [ponytail](https://ai-powered.com/tools/ponytail): ponytail is an open-source AI tool for ai models and llm tooling.

---
Source: https://ai-powered.com/tools/rllm
