🌐 Project • 📄 Paper • 💻 LLM-in-Sandbox-RL • 🤗 Huggingface • 📦 Model & Data • 🎬 Youtube • 📽️ Slides • 🕶️ Awesome Computer-Use-Agent • 🦞 ClawGym•

As vibe coding becomes common and 🦞 OpenClaw draws widespread attention, we present a systematic study to show that placing an LLM inside a code sandbox with basic computer functionalities lets it significantly outperform standalone LLMs across chemistry, physics, math, biomedicine, long-context understanding, and instruction-following with no extra training. RL further amplifies the gains.

• 📈 Consistent improvements across diverse non-code domains
• 🧠 File system as long-term memory, up to 8× token savings
• 🐳 Docker isolation for security (vs. unrestricted setups like 🦞 OpenClaw)
• 🔌 Works with OpenAI, Anthropic, vLLM, SGLang, etc.

Feel free to open an issue if you have any questions or run into any problems. We'd be happy to help!

▶️ Click to watch the demo video

• [2026-04-30] Released ClawGym, a scalable framework for building effective claw agents, covering data collection, training infra, and evaluation benchmark!
• [2026-04-11] Released scale-openclaw — collect 🦞 OpenClaw trajectories at scale on a single machine, Docker-free.
• [2026-03-25] Released Awesome Computer-Use Agents, a curated list tracking the development of computer-use agents across terminal and GUI paradigms, with an accompanying overview report.
• [2026-03-20] Uploaded talk slides on "LLM-in-Sandbox: From Coding Agent to General Agent".
• [2026-02-12] Released code, model & data, and wandb log for LLM-in-Sandbox Reinforcement Learning at LLM-in-Sandbox-RL.
• [2026-02-11] v0.2.0: Added benchmark module for reproducing paper results, evaluating any LLM, and adding your own tasks.
• [2026-01-23] Released paper, project page, and code. Honored as #1 paper of the day on Hugging Face.

• Installation
• Quick Start
• More Examples
• Benchmark and Reproduction
• Citation

Requirements: Python 3.10+, Docker

Skip this if Docker is already installed.

curl -fsSL https://get.docker.com -o get-docker.sh && sh get-docker.sh
dockerd > /var/log/dockerd.log 2>&1 &

Or follow the official Docker docs.

pip install llm-in-sandbox==0.2.0

Or install from source:

git clone https://github.com/llm-in-sandbox/llm-in-sandbox.git
cd llm-in-sandbox
pip install -e .

Docker Image

The default Docker image (cdx123/llm-in-sandbox:v0.1) will be automatically pulled when you first run the agent. The first run may take a minute to download the image (~400MB), but subsequent runs will start instantly.

llm-in-sandbox build

LLM-in-Sandbox is backend-agnostic: it connects to any OpenAI-compatible endpoint, so no local GPU is required when using API services (e.g., OpenAI, Anthropic, or a hosted vLLM/SGLang server). Self-hosting a model is optional, and its hardware needs depend on the chosen model and serving framework.

• Tested on: Ubuntu 22.04 with Python 3.10 and Docker.
• Installation time: ~1 minute on a normal desktop computer (the one-time Docker image pull, ~400 MB, also takes about a minute on a typical connection).
• Demo runtime: the quick-start demo completes in ~1 minute on a normal desktop computer (excluding model inference latency, which depends on your chosen endpoint).
• Dependencies: all Python dependencies are pinned in pyproject.toml; the sandbox runtime is provided by the Docker image cdx123/llm-in-sandbox:v0.1, pulled automatically on first run.

LLM-in-Sandbox works with various LLM providers including OpenAI, Anthropic, and self-hosted servers (vLLM, SGLang, etc.).

llm-in-sandbox run \
    --query "write a hello world in python" \
    --llm_name "openai/gpt-5" \
    --llm_base_url "http://your-api-server/v1" \
    --api_key "your-api-key"

vllm serve Qwen/Qwen3-Coder-30B-A3B-Instruct \
    --served-model-name qwen3_coder \
    --enable-auto-tool-choice \
    --tool-call-parser qwen3_coder \
    --tensor-parallel-size 8  \
    --enable-prefix-caching

llm-in-sandbox run \
    --query "write a hello world in python" \
    --llm_name qwen3_coder \
    --llm_base_url "http://localhost:8000/v1"  \
    --temperature 0.7

python3 -m sglang.launch_server \
    --model-path "deepseek-ai/DeepSeek-V3.2" \
    --served-model-name "DeepSeek-V3.2" \
    --trust-remote-code \
    --tp-size 8 \
    --tool-call-parser deepseekv32 \
    --reasoning-parser deepseek-v3 \
    --host 0.0.0.0 \
    --port 5678

llm-in-sandbox run \
    --query "write a hello world in python" \
    --llm_name DeepSeek-V3.2 \
    --llm_base_url "http://0.0.0.0:5678/v1" \
    --extra_body '{"chat_template_kwargs": {"thinking": True}}'

Run llm-in-sandbox run --help for all available parameters.

Each run creates a timestamped folder:

output/2026-01-16_14-30-00/
├── files/
│   ├── answer.txt      # Final answer
│   └── hello_world.py  # Output file
└── trajectory.json     # Execution history

We provide examples across diverse non-coding domains: scientific reasoning, long-context understanding, instruction following, travel planning, video production, music composition, poster design, and more.

👉 See examples/README.md for the full list.

Reproduce our paper results, evaluate any LLM in the sandbox, or add your own tasks.

👉 See llm_in_sandbox/benchmark/README.md

Feel free to open an issue if you have any questions or run into any problems, we’d be happy to help! You can also reach us directly at daixuancheng6@gmail.com and shaohanh@microsoft.com.

We learned the design and reused code from R2E-Gym. Thanks for the great work!

Apache License 2.0. See LICENSE for details.

If you find our work helpful, please cite us:

@article{cheng2026llm,
  title={Llm-in-sandbox elicits general agentic intelligence},
  author={Cheng, Daixuan and Huang, Shaohan and Gu, Yuxian and Song, Huatong and Chen, Guoxin and Dong, Li and Zhao, Wayne Xin and Wen, Ji-Rong and Wei, Furu},
  journal={arXiv preprint arXiv:2601.16206},
  year={2026}
}