4 independent teams built entirely different projects named MetaClaw — self-evolving memory, metagenomics analysis, agent scaffolding, and sandboxed execution. They share nothing but the name.

Meta-layer over OpenClaw / NanoBot / PicoClaw
Self-learning agent that evolves from every conversation. LoRA-based continual learning plus reinforcement learning, with no GPU cluster required.
Best for: Anyone who wants their agent to improve over time from its own usage. By far the most-starred MetaClaw and the one most people mean by the name.
| MetaClaw | MetaClaw | metaclaw | metaclaw | |
|---|---|---|---|---|
| Stars | 3.5K | 1 | 32 | 11 |
| Language | Python | Python | Shell | Go |
| Domain | Agent self-improvement | Metagenomics & multi-omics | Agent scaffolding | Agent runtime & sandboxing |
| Core mechanism | LoRA continual learning + RL | FlowHub upstream + 35 containerized skills | One-prompt project generation | Daemonless CLI + ClawCapsule artifacts |
| Publication | arXiv 2603.17187 | bioRxiv 2026.07.21.739769 | — | — |
| License | MIT | MIT | Unlicensed | Unlicensed |
| Best For | Agents that learn from usage | Auditable omics analysis | Fast multi-agent setup | Isolated reproducible execution |