SkillClaw: Collective Skill Evolution Framework
Let AI agents evolve skills from real interactions—across sessions, agents, devices, and users. Install and forget; evolution happens in the background.
Evidence shows: client proxy intercepts requests and records sessions, but least-privilege design not explicitly stated; user confirmation only in some commands (e.g., restore), not comprehensive; data flow transparency described but lacks detailed logging; sensitive data handling (e.g., API keys) configured but encryption not mentioned; dependency security no vulnerability scanning; external effects (e.g., modifying Hermes config) documented but no undo mechanism; rollback has restore command but no version rollback; source attribution has Acknowledgement but third-party licenses not clarified. Deductions: most trust criteria only partially met, lacking in-depth evidence.
Evidence shows: README and code structure consistent, tests cover protocol conversion, but no runtime verification; dependencies declared but not pinned, availability risk; failure messages not documented. Deductions: dependencies not pinned, failure handling documentation insufficient.
Evidence shows: clearly targets single-user, multi-agent, team scenarios, provides multiple install paths; capability boundaries described (client/server separation); trigger precision partially clear (e.g., command triggers), but auto-evolution trigger conditions not detailed; environment fit covers macOS/Linux/Windows, but Windows manual install. Deductions: auto-trigger details insufficient.
Evidence shows: README structure clear, good information architecture; install instructions detailed; naming stable but version not prominent; few examples, no FAQ; known limitations not listed; MIT license clear; version history in News but no formal CHANGELOG; maintenance responsibility not explicit. Deductions: missing FAQ and known limitations, informal version records.
Evidence shows: output is skill files, usability unverified; marginal value clearly described (skill evolution); cost-benefit not quantified. Deductions: lack of runtime verification and cost data.
Evidence shows: README claims support for multiple agents, but no verification evidence; tests cover some functionality, but not independently verified; facts and inferences not clearly separated. Deductions: claims lack traceability, tests not independently run.
- Dependencies are not pinned, posing supply chain risk.
- Auto-evolution may modify external agent configs; use with caution.
- Sensitive data (e.g., API keys) storage encryption not specified.
- Lack of formal CHANGELOG and known limitations documentation.
What does this agent do, and when should you use it?
SkillClaw is an open-source framework that makes LLM agents progressively better by evolving reusable skills from real session data. It serves both single users (skills auto-deduplicated, improved, and verified across sessions) and teams, where multiple agents, devices, or users can feed the same evolution loop. The system comprises a local client proxy that intercepts agent requests and records session artifacts, and an optional evolve server that reads session data from shared storage, evolves or creates skills, and writes them back. The two components interact only through shared storage (local filesystem, Alibaba OSS, or S3). Skills are stored in a standardized SKILL.md format. The project is MIT-licensed and is accompanied by an arXiv paper.
SkillClaw operates as two components. The client proxy listens on /v1/chat/completions and /v1/messages, intercepting agent requests, recording session artifacts, and managing a local skill library. The evolve server (evolve_server) consumes session data from shared storage and runs either a workflow engine (fixed 3-stage LLM pipeline: Summarize → Aggregate → Execute) or an agent engine (OpenClaw-driven workspace with direct skill editing) to evolve or create skills, which are then written back to the shared skill store. Skills are stored as SKILL.md files, with storage backends for local filesystem, Alibaba OSS, and S3. Users interact via the skillclaw CLI, running commands like setup, start --daemon, status, skills pull/push/sync, and dashboard sync/serve for inspection. For Hermes, SkillClaw rewrites ~/.hermes/config.yaml to route through the local proxy and prepares ~/.hermes/skills.
- A solo developer using Hermes or an OpenAI-compatible API wants to automatically curate and improve their skill library without manual overhead.
- A team shares a skill store, so one member's debugging experience evolves a skill that others benefit from instantly.
- A user with multiple devices (home, school, work) wants a unified skill library that follows them, avoiding re-learning on each machine.
- A user running multiple agents (Hermes, Codex, Claude Code, OpenClaw, etc.) wants a single skill library that merges and cross-pollinates skills across agents.
- A group operates a shared evolve server that automatically evolves, validates, and publishes skills from sessions across the team.
- A team with quality gates enables validation mode, where client proxies validate candidate skills in the background before they are published.
What are this agent's strengths and limitations?
- Automatic skill evolution: continuously extracts, deduplicates, and refines skills from real interactions, reducing manual maintenance.
- Collective intelligence: supports multi-user, multi-agent, and multi-device skill sharing, amplifying team experience.
- Broad integration: natively supports Hermes, Codex, Claude Code, OpenClaw, and other OpenAI-compatible APIs.
- Flexible deployment: client and server decouple through local storage, OSS, or S3, allowing scalable setups.
- Depends on external LLM API providers and possibly paid storage (OSS/S3), incurring ongoing costs.
- Requires Python 3.10+ and some technical setup, which may deter non-technical users.
- The server's agent engine requires installing OpenClaw globally.
- Increased operational complexity: running a proxy and potentially a server adds moving parts.
- No native Windows installer; manual installation needed.
How do you install or deploy this agent?
Requires Python 3.10+, macOS/Linux/Windows, and an OpenAI-compatible API account or AWS Bedrock. On macOS/Linux: git clone https://github.com/AMAP-ML/SkillClaw.git && cd SkillClaw && bash scripts/install_skillclaw.sh && source .venv/bin/activate. On Windows PowerShell (manual): git clone ...; Set-Location SkillClaw; python -m venv .venv; .\.venv\Scripts\Activate.ps1; pip install -e ".[evolve,sharing,server]".
How do you use this agent?
Run skillclaw setup to configure provider, model, local skills directory, and optional shared storage. Start the client proxy with skillclaw start --daemon. For single-user auto-evolution, run skillclaw-evolve-server --use-skillclaw-config --interval 300 --port 8787. For team sharing, configure shared storage and start the evolve server. Manage skills with skillclaw skills pull/push/sync. Inspect with skillclaw dashboard sync/serve. Enable validation mode via skillclaw config validation.enabled true.
How does this agent compare with similar options?
Compared to similar projects like MetaClaw and OpenClaw-RL, SkillClaw emphasizes collective skill evolution with explicit support for cross-user and cross-agent sharing, and it ships a dedicated evolve server with validation workflows.