The best part of OpenClaw is the massive ecosystem of ClawHub skills, but the worst part is dedicating ~200MB of idle RAM to a Node runtime just to host a few agents. I noticed most lightweight remakes fail because they break compatibility with existing skills.

I built Capabot to solve this. It is a single-binary, pure Go runtime that parses and executes any OpenClaw SKILL.md out of the box, but drops the idle memory to ~23MB and cold starts in <100ms.

A few architectural differences from the original: - Skill Security: Instead of giving the LLM raw shell access, any skills that require real computation are compiled to WASM and run in a strict, isolated sandbox using wazero. - Automations/Cron: Scheduling background agent prompts using Node's event loop is heavy. Capabot uses native goroutines, making scheduled cron agents incredibly efficient. - Transports: Natively holds WebSocket connections open for Discord, Telegram, and Slack without the Node overhead.

It still has all the other key features of OpenClaw. The embedded React UI makes it practically frictionless to self-host on a cheap VPS. It currently supports Anthropic, OpenAI, Gemini, and OpenRouter.

I'd love to hear your thoughts on the wazero sandboxing or the ReAct loop implementation! Fully open-source under MIT liscence.