SuperAPI AI Directory
Partner with us中文AI newsFavorites
DirectoryNews and tutorials

SkillClaw - AMAP's Open-Source Framework for Collective Evolution of AI Agent Skills

Author: AI小集 · Original source: AI工具集

What is SkillClaw

SkillClaw is an open-source framework for collective evolution of AI Agent skills from AMAP-ML (Amap Machine Learning Team). The framework intercepts Agent sessions through a local API proxy, and in the background automatically extracts, deduplicates, optimizes, and archives reusable skills from real interactions, enabling cross-device, cross-Agent, and cross-team skill sharing and continuous evolution. Users only need to converse with the Agent as usual, with zero additional operations throughout the skill evolution process.

SkillClaw

Main Features of SkillClaw

  • Zero-intervention automatic evolution: Adopts a dual-loop architecture of task-time loop + evolution loop. After tasks are completed, skills are automatically distilled in the background, fully transparent to users.
  • Broad cross-framework compatibility: Natively supports mainstream Agent frameworks such as Hermes, Claude Code, OpenClaw, Codex, QwenPaw, IronClaw, PicoClaw, and any OpenAI-compatible API.
  • Collective skill sharing: The experience of any team member can be automatically accumulated into shared skills, directly benefiting other members and avoiding repeated pitfalls.
  • Autonomous skill quality: Automatically deduplicates and merges similar skills, optimizes low-quality skills, and archives outdated skills, preventing skill library bloat and confusion.
  • Bilingual visual dashboard: Supports skillclaw dashboard command to launch a local panel for viewing skill version comparisons, verification status, and session traceability.

Technical Principles of SkillClaw

  • Client Proxy: Local API proxy (intercepts /v1/chat/completions and other endpoints), records session data, and synchronizes skills with shared storage.
  • Workflow Evolve Server: A fixed three-stage LLM pipeline: Summarize (summarize sessions) → Aggregate (identify duplicate patterns) → Execute (evolve or create skills and write back).
  • Agent Evolve Server: An alternative solution based on OpenClaw that can autonomously read sessions, analyze trajectories, and directly edit skill files.

How to Use SkillClaw

  • Clone and install: Run git clone https://github.com/AMAP-ML/SkillClaw.git and run the installation script to create a local Python virtual environment.
  • Initialize configuration: Run skillclaw setup, and follow the wizard to select the Agent framework, configure the upstream API key and storage backend (local or OSS/S3).
  • Start the proxy: Run skillclaw start --daemon to start the local API proxy, which automatically intercepts and records Agent session data.
  • Converse normally: Interact with the Agent to complete tasks as usual. SkillClaw silently captures sessions in the background, with no additional operations required.
  • Deploy the evolution server(optional): Run skillclaw-evolve-server to start the background evolution service, which automatically distills, optimizes, and archives skills from sessions.
  • Synchronize the skill library: Use skillclaw skills pull/push/sync to bidirectionally synchronize skills between local and shared storage, enabling team sharing.
  • View the dashboard(optional): Run skillclaw dashboard serve to launch the local visual panel for viewing skill version comparisons and evolution trajectories.

Key Information and Usage Requirements of SkillClaw

  • Product Name：SkillClaw
  • Development Team: AMAP-ML (Amap Machine Learning Team)
  • Open Source License: Open source (GitHub 1.2K+ Stars)
  • System Requirements：macOS / Linux / Windows，Python ≥ 3.10
  • API Requirements: OpenAI-compatible API or AWS Bedrock
  • Deployment Mode: Personal local / team shared storage (OSS/S3/local)
  • Installation Method：git clone + bash scripts/install_skillclaw.sh
  • Startup Command：skillclaw setup → skillclaw start --daemon
  • Optional Dependencies：openclaw(only required for the Agent evolution engine)

Core Advantages of SkillClaw

  • Truly Zero Intrusion: Does not change existing workflows; just install it and chat normally, and skill evolution happens silently.
  • Compounding Experience: Experience from multiple personal devices and multiple Agents is consolidated in a unified way; during team collaboration, one person hits a pitfall and everyone becomes immune.
  • No Framework Lock-in: Not tied to a specific Agent ecosystem, covering almost all current mainstream CLI Agents.
  • Flexible Storage: Supports smooth scaling from purely local single-machine use to OSS/S3 team sharing, with no need to refactor the architecture.
  • Controllable Quality: Built-in validation mechanism (Validation Worker); candidate skills must pass scoring and approval before being officially released.

SkillClaw Project Address

  • GitHub Repository：https://github.com/AMAP-ML/SkillClaw

Comparison of SkillClaw with Similar Competing Products

Dimension SkillClaw Mem0 Letta (formerly MemGPT)
Positioning Agent skill collective evolution framework General AI Agent memory layer Agent runtime + OS-style memory management
Skill Evolution Automatically distilled from real conversations SKILL.md, automatically deduplicated, optimized, and archived Supports Procedural Memory to store processes, passively extracts facts The Agent autonomously edits three layers of memory, with no automatic skill distillation and archiving
Sharing Mechanism OSS / S3 / local shared storage, native team support Cloud API + OpenMemory MCP local version, cross-application memory sharing Self-hosted stateful service; teams need to deploy sharing themselves
Framework Compatibility Plug-and-play across multiple frameworks such as Hermes, Codex, Claude Code, OpenClaw, etc. Any Agent can connect via SDK / REST / MCP Requires using the Letta runtime; the Agent loop is deeply bound to memory, resulting in high migration costs
Deduplication and Optimization Automatically merges similar skills, eliminates low-quality skills, and keeps the skill library clean Relies on external entity resolution and reflection mechanisms for deduplication, with no native skill governance No automatic deduplication; memory quality depends entirely on the Agent's autonomous judgment
Deployment Method Local proxy + optional evolution server, progressive scaling Managed SaaS or open-source self-hosting, lightweight integration Self-hosted stateful service, relatively high operational complexity

Application Scenarios of SkillClaw

  • Individual Developers: Use Agents on multiple devices at home/company/school, with the skill library synchronized in real time and no need for repeated training.
  • Multi-Agent Collaboration: React patterns learned by frontend Agents are automatically shared with backend Agents, improving API design efficiency.
  • Team knowledge accumulation: Technical teams automatically solidify debugging experience and code standards into reusable skills, reducing the onboarding cost for new members.
  • Automated skill library operations: Replaces manual organization of .hermes/skills and other directories, solving the problems of duplicate, outdated, and disorganized skills.