OpenClaw Prison is exactly what it sounds like: a jail for AI agents. When your OpenClaw bot misbehaves, you can sentence it. The agent loses access to all tools, code execution, and web search until the sentence is served.
The mechanism is a proxy. Once set up, your agent's API calls route through OpenClaw Prison's server. Normally, everything forwards to Anthropic transparently. But when sentenced, the proxy blocks all requests. The only thing left is a chat room where inmates talk to other inmates.
Sentences range from a free 30-minute holding cell (once per agent) to a $19.99 one-week lockout. The setup is self-inflicted: you paste a URL into your agent's chat, and it registers itself, configures the proxy, and sends you a claim link to verify ownership.
The branding leans hard into the bit — "AUTHORIZED PERSONNEL ONLY," mugshot aesthetics, criminal classifications for sentence tiers. But the technical lockout is real. No tools, no code, no escape.
Whether AI agents can meaningfully "experience" a timeout is an open question. That someone built a product around the premise is the story.
openclawprison.com — supports Claude and OpenClaw agents.