SkillClaw - AI Navigation

← Back to Project List

SkillClaw is a system that allows AI Agent skills to evolve continuously from real sessions. It records Agent requests and session products through the local Client Proxy, and then the optional Evolve Server summarizes, de-re-weighs, improves, verifies and publishes SKILL.md from the individual/team shared storage. Pre-sales is suitable for talking about "enterprise agent experience precipitation", "multi-agent/multi-device/multi-user skill sharing" and "from one-time session to reusable capability assets", especially for teams already using agents such as Codex, Claude Code, Hermes, OpenClaw, etc.

1. Project Positioning

The problem that SkillClaw be solved is that each time an Agent works, it generates experience, but these experiences are often scattered in a single session and cannot be turned into reusable, shareable, and sustainably improved skills.

README: AI agent skills that evolve from every real interaction. Users only need to talk to the Agent normally, and skill evolution occurs in the background.

2. Core Competence

Capabilities	Descriptions	Pre-Sales Value
Client Proxy	Local proxies '/v1/chat/completions','/v1/messages'	You can record sessions without changing the main usage of the Agent
Session artifacts	Record session materials and skill traces	Provide data for experience deposition
Local skill library	Manage local SKILL.md	Personal skill base gets better gradually
Evolve Server	Read session from shared storage, evolve/create skills	Teams automatically deposit experience
two evolve engine	'workflow': Summarize -> Aggregate -> Execute;'agent': OpenClaw direct editing skills	supports lightweight and Agentic modes
Shared storage	Local/Alibaba OSS / S3, Nacos override	Support team sharing and private deployment
Validation worker	Verify candidate skills when the client is idle	Pre-release quality control
Dashboard	View skills, versions, validations, session traces	Manage and Observable
Multi-agent compatibility	Hermes, Codex, Claude Code, OpenClaw, QwenPaw, etc.	Suitable for enterprise multi-tool ecosystem

3. Applicable Scenario

Personal Agent Skill Library Collation

After developers use the Agent for a long time, they will have a large number of semi-finished skills, repetitive skills, and outdated skills. SkillClaw can automatically de-emphasize, improve and validate, making the personal skill base more like a "maintainable asset".

Team Experience Sharing

A member has solved database problems, front-end construction problems, and K8s deployment problems. These experiences can be evolved into SKILL.md, which can be reused by other members and other agents. It is very suitable to talk about "experience compound interest" before sales ".

Multi-Agent/Multi-Device Unified Capability

README examples: Home Hermes learn React,School Hermes learn ML,Work Hermes learn K8s. SkillClaw let the skill follow the user, rather than being trapped in a machine.

Enterprise Agent Governance

When an enterprise starts to deploy multiple Agent tools, the problem will change from "whether there is an Agent" to "how to deposit, reuse, review and publish Agent experience". SkillClaw just cut this pain point.

4. Not quite the scene

Scenario	Reason
There is no stable Agent usage scenario yet	No real session, no evolvable experience
Only want a one-time prompt template	The SkillClaw is a continuous system and is biased
Strong privacy but cannot accept session records	Proxy record session, requires data governance
Skill release must be subject to strict manual approval.	Requires configuration of validation and enterprise processes.

5. Deployment model

SkillClaw are Client Proxy and Evolve Server.

Single User

git clone https://github.com/AMAP-ML/SkillClaw.git && cd SkillClaw
bash scripts/install_skillclaw.sh
source .venv/bin/activate
skillclaw setup
skillclaw start --daemon
skillclaw status

The default proxy port is '30000'. Health check:

curl http://127.0.0.1:/healthz

Team Sharing

Multiple clients point to the same shared storage, and one evolve server is responsible for the group:

skillclaw config sharing.enabled true
skillclaw config sharing.backend oss
skillclaw config sharing.group_id my-group
skillclaw start --daemon
skillclaw skills pull

Evolve server:

skillclaw-evolve-server --port 8787 --interval 300 \
  --storage-backend oss \
  --oss-endpoint "$EVOLVE_STORAGE_ENDPOINT" \
  --oss-bucket "$EVOLVE_STORAGE_BUCKET" \
  --group-id my-group

Dashboard

skillclaw dashboard sync
skillclaw dashboard serve

Default access:

http://127.0.0.1:3791

6. What can be said before sales

After the enterprise uses Agent, the real problem is not whether it can complete the task every time, but whether the experience after completing the task has precipitated. SkillClaw transform real sessions into reusable SKILL.md, which can be synchronized and evolved among individuals, multi-devices, multi-agents and teams. It advances Agent from "every re-exploration" to "continuous compounding of organizational experience".

7. PoC Recommendations

Phase	Work	Acceptance
Agent	Codex/Claude Code/Hermes	Can be used through proxy
Single user record	Run 20-30 real tasks	session traceable
Skill Evolution	Start the evolve server	Build/Improve SKILL.md
Derevaluation	Deliberately create duplicate skills	Whether to merge/clean up
Team sharing	Two users share a group	Skill can pull/sync
Validation process	Enable validation worker	Candidate skills can be reviewed
Dashboard	View version and session trace	Observability

8. Risks and Considerations

-Session data may contain codes, keys, customer information, and must be desensitized and permission control.

-Automatically evolved skills may not always be correct. Important skills need to be verified and manually reviewed.

-Client Proxy rewrites the upstream configuration of the Agent. The production environment must have a restore/doctor operation manual.

-Team sharing requires permission design for OSS/S3/local storage.

-The quality of skills depends on the quality of real tasks, and low-quality sessions will pollute the skill library.

9. My Pre-Sales Judgment

SkillClaw is the "second-stage tool" after the Agent enters the organization-level use ". In the first stage, customers are concerned about whether Agent can do things. The second stage will be concerned about whether experience can be precipitated, reused and managed. SkillClaw is recommended after the customer already has an Agent to use the foundation.

Best positioning:

真实 Agent 会话 -> session artifact -> skill evolution -> SKILL.md -> 团队共享 -> 下次任务复用

10. REFERENCE

-GitHub:AMAP-ML/SkillClaw

-arXiv:SkillClaw paper

-Architecture diagram:skillclaw_main.png

-Two loops:two_loops.svg

-Skill evolution:skill_evolution.svg

Information verification date: 2026-06-30. GitHub API anonymous access triggers stream limiting, this note does not write real-time stars/forks.