SkillClaw - Amap's open-source AI Agent skills collective evolution framework - AiBoss

What is SkillClaw?

SkillClaw is an open-source AI Agent skills collective evolution framework developed by AMAP-ML (Gaode Map's machine learning team). The framework intercepts Agent sessions through a local API proxy, automatically extracting, deduplicating, optimizing, and archiving reusable skills from real interactions in the background. This enables skill sharing and continuous evolution across devices, Agents, and teams. Users only need to interact normally with the Agent; the entire skill evolution process requires zero additional steps.

SkillClaw's main functions

Zero-intervention automatic evolutionIt adopts a dual-loop architecture of task time loop + evolution loop. After the task is completed, the background automatically extracts skills, which is completely transparent to the user.
Broad cross-framework compatibilityNative support for mainstream agent frameworks such as Hermes, Claude Code, OpenClaw, Codex, QwenPaw, IronClaw, and PicoClaw, as well as any OpenAI-compatible API.
Collective Skills SharingThe experience of any member of the team can be automatically recorded as shared skills, directly benefiting other members and avoiding repeating the same mistakes.
Skills quality autonomyAutomatically deduplicate and merge similar skills, optimize low-quality skills, and archive outdated skills to prevent the skill library from becoming bloated and chaotic.
Bilingual Visual Dashboard:support skillclaw dashboard The command launches the local panel, allowing you to view skill version comparisons, verification status, and session traceability.

SkillClaw's technical principles

Client ProxyLocal API Proxy (Interception) /v1/chat/completions (Endpoints, etc.) record session data and synchronize with shared storage.
Workflow Evolve ServerThe fixed three-stage LLM pipeline is: Summarize (summarize the session) → Aggregate (identify recurring patterns) → Execute (evolve or create skills and write them back).
Agent Evolve ServerAn alternative based on OpenClaw that can autonomously read sessions, analyze trajectories, and directly edit skill files.

How to use SkillClaw

Clone and install:implement git clone https://github.com/AMAP-ML/SkillClaw.git Then run the installation script to create a local Python virtual environment.
Initialization configuration:run skillclaw setupFollow the wizard to select the Agent framework, configure the upstream API key and storage backend (local or OSS/S3).
Start Agent:implement skillclaw start --daemon Start the local API agent to automatically intercept and record Agent session data.
Normal conversationInteract with the Agent as usual to complete tasks. SkillClaw silently captures the session in the background without any additional operations.
Deploy Evolution Server(Optional): Run skillclaw-evolve-server Start the background evolution service to automatically extract, optimize, and archive skills from conversations.
Synchronized Skills Library:use skillclaw skills pull/push/sync The ability to synchronize skills between local and shared storage enables team sharing.
Check the dashboard(Optional): Execute skillclaw dashboard serve Launch the local visualization panel to view skill version comparisons and evolution trajectories.

Key information and usage requirements of SkillClaw

Product NameSkillClaw
Development TeamAMAP-ML (Gaode Map Machine Learning Team)
Open source licenseOpen source (GitHub 1.2K+ Stars)
System Requirements: macOS/Linux/Windows, Python ≥ 3.10
API Requirements: OpenAI-compatible API or AWS Bedrock
Deployment modePersonal 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(Required only for Agent Evolution Engine)

SkillClaw's core advantages

Truly zero intrusionWithout changing the existing workflow, you can chat normally after installation. Skill evolution will be completed silently.
Experience can be compoundedExperience gained from working with multiple devices and agents can be consolidated and consolidated; when one person makes a mistake in a team collaboration, the entire team is immune to the problem.
Frame unlockIt is not bound to a specific agent ecosystem and covers almost all current mainstream CLI agents.
Flexible storageSupports smooth scaling from purely local standalone machines to OSS/S3 team sharing without requiring architecture refactoring.
Quality controllableThe built-in validation mechanism (Validation Worker) requires candidate skills to be scored and approved before they are officially released.

SkillClaw's project address

GitHub repositoryhttps://github.com/AMAP-ML/SkillClaw

SkillClaw's Competitive Product Comparison

Dimension	SkillClaw	Mem0	Letta (formerly MemGPT)
position	Agent Skill Collective Evolution Framework	General AI Agent Memory Layer	Agent runtime + OS-style memory management
Skill Evolution	Automatically extracting from real conversations `SKILL.md`Automatic deduplication, optimization, and archiving	Supports procedural memory storage processes, passively retrieving facts.	The Agent autonomously edits three layers of memory, but lacks automatic skill refinement and archiving.
Sharing mechanism	OSS / S3 / Local Shared Storage, natively supported by the team	Cloud API + OpenMemory MCP local version, enabling cross-application memory sharing.	Self-hosted stateful services require teams to deploy and share them themselves.
Framework compatibility	Plug and play with multiple frameworks such as Hermes, Codex, Claude Code, and OpenClaw.	Any Agent can connect via SDK / REST / MCP	The Letta runtime is required, and the Agent loop is deeply bound to memory, resulting in high migration costs.
Deduplication optimization	Automatically merge similar skills, eliminate low-quality skills, and keep the skill library clean.	Deduplication relies on external entity parsing and reflection mechanisms, lacking native skill-based governance.	Without automatic deduplication, the memory quality relies entirely on the Agent's autonomous judgment.
Deployment method	Local proxy + optional evolution server, progressive scaling	Managed SaaS or open-source self-hosted solutions, lightweight integration.	Self-managed stateful services have high operational complexity.

Application Scenarios of SkillClaw

individual developersUse the Agent on multiple devices at home/office/school, with the skill library synchronized in real time, eliminating the need for repetitive training.
Multi-agent collaborationThe React patterns learned by the front-end agent are automatically shared with the back-end agent, improving API design efficiency.
Team knowledge accumulationThe technical team automatically solidifies debugging experience and coding standards into reusable skills, reducing the learning curve for newcomers.
Automated Skills Library Operations and Maintenance: Replacement for manual sorting .hermes/skills The catalog addresses issues such as repetitive, outdated, and disorganized skills.