The Proxy That Evolves Agent Skills in the Background

What it does SkillClaw runs as a local API proxy that intercepts your agent’s chat completions and messages. It records session artifacts to a local or shared store, then distills them into evolving SKILL.md files that are deduplicated, improved, and synced back to your agents. An optional background evolve server handles the refinement, so the learning happens after the conversation ends.

The interesting bit The whole thing is transparent: you keep talking to your agent normally while the proxy silently rewrites its growth curve. Because storage is decoupled—using S3, Alibaba OSS, or the local filesystem—the same skill library can follow you across devices or feed a whole team’s agents. It’s institutional knowledge, except it auto-merges and auto-updates.

Key highlights

• Compatible with Hermes, Codex, Claude Code, OpenClaw, QwenPaw, IronClaw, and any OpenAI-compatible API
• Two evolution engines: a fixed 3-stage LLM pipeline (Summarize → Aggregate → Execute) or an OpenClaw-driven agent workspace
• Shared storage layer lets solo users, multi-device setups, and teams all feed the same evolution loop
• Includes a bilingual dashboard for inspecting skill libraries, validation progress, and session traces
• Auto-deduplication and cross-pollination: a frontend agent’s React patterns can improve a backend agent’s API design

Caveats

• Windows installation is manual; the automated shell installer only covers macOS and Linux
• The beginner path has been locally smoke-tested on macOS, so other platforms may need patience
• Automatic background evolution requires running and configuring a separate evolve_server component

Verdict A solid bet for developers running multiple agents who want shared, self-improving muscle memory. Probably overkill if you just want a single chatbot wrapper without backend plumbing.