pome223/boiled-claw-public
Reference architecture for closed-loop AI agents — plan, execute, verify, repair.
Platform-specific configuration:
{
"mcpServers": {
"boiled-claw-public": {
"command": "npx",
"args": [
"-y",
"boiled-claw-public"
]
}
}
}Add the config above to .claude/settings.json under the mcpServers key.
<p align="center"> </p>
<h1 align="center">boiled-claw</h1>
<p align="center"> A reference architecture for closed-loop AI agents — plan, execute, verify, repair. </p>
Agents without a verification loop can't be trusted in production. boiled-claw is an AI agent architecture with a closed-loop execution cycle: Planner → PolicyJudge → Executor → Verifier → Repair.
Built on Google Agent Development Kit (ADK). MIT License. Fork-friendly, upstream-curated.
Homepage: astropomeai.com
> [!WARNING] > boiled-claw can execute shell commands, read and write files, control browsers, and drive desktop UI on the host machine. > This repository is a reference implementation, not a hardened security product. Tool approvals and security policies reduce risk, but do not guarantee safety. > Do not expose Gateway / Host Bridge / Desktop Bridge to untrusted networks. Run it only in environments you control, and prefer isolated or disposable machines for experimentation. > By using or running this code, you accept all risk. To the maximum extent permitted by law, the author disclaims all liability for any damage, data loss, security incident, account action, system instability, or other harm resulting from its use, misuse, or modification.
> Note: This is a maintainer-led reference implementation. Upstream is curated for design coherence — no support or review commitment. See CONTRIBUTING.md.
Start with ARCHITECTURE.md if you want the design rationale before the implementation details.
Loading reviews...