ClawGym is a scalable framework for synthesizing data, training agents, and evaluating Claw-style personal agents in local, stateful workspaces. 👨💻
⚠️ Code, datasets, benchmark assets, and model checkpoints are currently under internal company review and will be released soon. 🙏
- [2026.05.15] 🔥 We release ClawGym-Bench, including the benchmark data and evaluation script.
- [2026.05.15] 🔥 We release ClawGym-Agents, including our 13.5K synthesized tasks and 24.5K training trajectories, with corresponding SFT and RL training code.
- [2026.05.15] 🔥 We release our SFT checkpints: ClawGym-4B, ClawGym-8B, and ClawGym-30A3.
- [2026.05.01] 🚀 We release Claw Series Part I: ClawGym, our first work on scalable data synthesis, training, and evaluation for Claw-style agents.
ClawGym is a scalable framework for building, training, and evaluating Claw-style personal agents across realistic local workspace environments.
ClawGym supports the full lifecycle of personal-agent development. It first constructs ClawGym-SynData, a diverse dataset of 13.5K executable tasks synthesized from persona-driven intents and skill-grounded operations. Each task is paired with a realistic mock workspace and hybrid verification mechanisms, enabling reproducible execution and automated evaluation.
Based on these synthesized tasks, we collect interaction trajectories through black-box rollouts and use them to train a family of capable Claw-style models, termed ClawGym-Agents. We further explore reinforcement learning (RL) through a lightweight sandbox-parallel pipeline, supports both Docker-based and Docker-free backends, and learns directly from outcome rewards.
To support reliable evaluation, we build ClawGym-Bench, a benchmark of 200 instances calibrated through automated filtering and human-LLM review.
ClawGym-SynData contains 13.5K executable Claw-style tasks. It combines two synthesis routes:
- Persona-driven synthesis: samples user profiles, scenario categories, and atomic operations to generate realistic workspace-grounded requests.
- Skill-grounded synthesis: builds tasks from OpenClaw skills, using one primary skill with optional supporting skills to encourage multi-step workflows.
The task generation process covers 9 scenario categories, 43 subcategories, 7 operation categories, and 26 atomic operations. For skill-grounded synthesis, we annotate 16,837 collected skills across categories such as Data & APIs, Dev Tools, Workflows, Automation, Security, Prompts, MCP Tools, and others.
Each task is paired with lightweight mock resources and task-specific verifiers. Human-sampled quality analysis over 50 training tasks gives an overall score of 4.06 / 5, indicating good task coherence, resource consistency, and verifier quality.
ClawGym-Agents are trained from black-box OpenClaw rollouts on ClawGym-SynData. We collect 24.5K interaction trajectories using teacher rollouts from MiniMax-M2.5 and GLM-5.1, then filter trajectories by verifier scores.
The selected trajectories are long-horizon and tool-intensive:
| Avg. Rounds | Avg. Tokens | Avg. Tool Calls | Avg. Tool Types |
|---|---|---|---|
| 13.00 | 18.67K | 15.82 | 3.25 |
We perform multi-turn SFT on Qwen3-series backbones and obtain ClawGym-4B, ClawGym-8B, and ClawGym-30B-A3B. We also explore reinforcement learning (RL) through a lightweight sandbox-parallel pipeline.
ClawGym-Bench is a diagnostic benchmark of 200 instances for Claw-style agents. Each task contains a user instruction, mock workspace resources, and a task-specific verifier.
- 156 tasks use code-based verification.
- 44 tasks use hybrid verification, combining code checks with rubric-based judgment.
- Hybrid scoring uses 0.7 weight for code-based verification and 0.3 weight for rubric-based verification.
The benchmark is selected through difficulty-aware filtering and human-LLM review. It covers six workspace-grounded categories:
| Category | Product. & Collab. |
Systems & Auto. |
Analysis & Reason. |
Content & Domain |
Planning & Knowl. |
Software Dev. |
|---|---|---|---|---|---|---|
| # Tasks | 44 | 42 | 35 | 28 | 26 | 25 |
We evaluate ClawGym-Agents on ClawGym-Bench and PinchBench. The main results show that training on ClawGym-SynData consistently improves compact open-weight backbones.
Our goal is to provide a transparent and reproducible foundation for building and evaluating Claw-style personal agents. We plan to release the code and data in several stages:
- ✅ ClawGym-Bench data
- ✅ Evaluation code
- ✅ ClawGym-Agents checkpoints
- ✅ Training data
- ✅ SFT code
- ✅ RL code
- ⬜ Data synthesis pipeline code
Our implementation builds upon the excellent codebases of slime, OpenClaw, OpenClaw-RL, PinchBench, OpenRLHF and Megatron-LM.
We sincerely thank these projects for their valuable insights and high-quality implementations, which have greatly facilitated our research.
@article{bai2026clawgym,
title={ClawGym: A Scalable Framework for Building Effective Claw Agents},
author={Bai, Fei and Song, Huatong and Sun, Shuang and Cheng, Daixuan and Yang, Yike and Hao, Chuan and Li, Renyuan and Chang, Feng and Wei, Yuan and Tao, Ran and others},
journal={arXiv preprint arXiv:2604.26904},
year={2026}
}