AMAP-ML/SkillClaw

SkillClaw – Collective Skill Evolution for LLM‑Agents

What it is – SkillClaw is an open‑source system that sits between an LLM‑based agent (e.g., Hermes, Codex, Claude Code, QwenPaw, OpenClaw, etc.) and the LLM API. It records every interaction, extracts skills (re‑usable snippets of code, prompts, or knowledge), deduplicates them, and continuously improves them. The result is a shared skill library that grows automatically as you and your teammates use the agents.

Why it matters – Modern LLM agents often accumulate ad‑hoc “knowledge” that never gets reused. SkillClaw turns that accidental learning into a structured, versioned skill store that can be shared across:

• multiple agents on the same machine,
• the same agent on different devices, and
• whole teams working on a common project.

Core components

Component	Role
Client proxy (skillclaw)	A local HTTP proxy that mimics the OpenAI /v1/chat/completions endpoint. It intercepts requests, logs session artifacts, and serves the current skill library to the agent.
Evolve server (skillclaw‑evolve‑server)	Optional background service that reads the logged sessions, runs a three‑stage LLM pipeline (Summarize → Aggregate → Execute) or an OpenClaw‑driven agent workspace, and writes back updated SKILL.md files.
Storage layer	Filesystem, Alibaba OSS, or S3. Both client and server only communicate through this shared storage, making deployment flexible (local only, remote VM, or cloud bucket).

How you use it

1. Install – One‑line installer for macOS/Linux (bash scripts/install_skillclaw.sh) or a manual pip -e install for Windows.
2. Configure – Run skillclaw setup; the wizard asks for your LLM provider, model, where to store skills, and (optionally) which agent to auto‑configure (Hermes, Codex, etc.).
3. Start the proxy – skillclaw start --daemon. The proxy listens on port 30000 by default and rewrites the agent’s config to point at http://127.0.0.1:30000.
4. Chat as usual – Your agent talks to the LLM through the proxy; SkillClaw silently records the session and updates the local skill directory.
5. (Optional) Run the evolve server – skillclaw‑evolve‑server --use-skillclaw-config … to automatically turn recorded sessions into refined skills. For teams, run a single server that all clients point at the same OSS/S3 bucket.

Key features

• Zero‑effort evolution – No extra commands after the initial setup; skills improve in the background.
• Auto‑deduplication & versioning – Identical or overlapping skills are merged, and each version is stored as SKILL.md with a changelog.
• Broad agent compatibility – Native integrations for Hermes, OpenClaw, QwenPaw, IronClaw, PicoClaw, ZeroClaw, Codex, Claude Code, and any OpenAI‑compatible API.
• Cross‑device & cross‑user sharing – Store the skill library in OSS/S3 (or a shared local folder) so the same knowledge follows you from home to work and can be pooled across a team.
• Dashboard – skillclaw dashboard sync/serve provides a local web UI to browse skills, see version history, and inspect session traces.
• Validation worker – Optional background validator that reviews candidate skills before they are published to the shared store.
• Extensible engines – workflow (fixed LLM pipeline) or agent (OpenClaw‑driven workspace) for skill generation.

Typical workflows

• Solo developer: Install client, run proxy, optionally start a local evolve server. Your personal skill library becomes smarter with each chat.
• Team of engineers: One evolve server runs on a cloud VM, all members point their clients at the same OSS bucket. When Alice solves a deployment bug, the generated skill is instantly available to Bob, Carol, etc.
• Multi‑agent setup: Run Hermes for frontend UI, Codex for code generation, and QwenPaw for data‑analysis; SkillClaw merges the skills each agent discovers, letting the frontend benefit from backend insights and vice‑versa.

Getting started (quick‑start for a single user on macOS/Linux)

git clone https://github.com/AMAP-ML/SkillClaw.git && cd SkillClaw
bash scripts/install_skillclaw.sh
source .venv/bin/activate
skillclaw setup          # answer prompts (default values work for a test run)
skillclaw start --daemon
# Verify proxy health
curl http://127.0.0.1:30000/healthz

Now run any supported agent (e.g., hermes chat -m skillclaw-model …) and watch the skill library grow under ~/.hermes/skills.

Documentation & community

• Paper: arXiv 2604.08377 (PDF & Hugging Face summary linked in the badge bar).
• Chinese README: assets/README_ZH.md.
• WeChat group for real‑time discussion (badge link).
• News section in the README tracks recent releases (e.g., dashboard, Codex/Claude integration, LongHorizon‑Harness).

License – MIT, so you can freely embed SkillClaw in commercial or research projects.

---

All information above is taken directly from the repository’s README; no external assumptions have been added.