openclaw-skill-manager

An OpenClaw agent skill that gives agents autonomous skill-management capabilities. When the agent identifies a capability gap, it can create a new skill, push it to GitHub, and install it after user approval.

What it does

The skill-manager enables a self-improving workflow:

1. Detect — Agent identifies it needs a capability it doesn't have
2. Create — Agent generates a new skill (SKILL.md + tools), pushes to GitHub
3. Track — New skill is recorded as "pending" (not installed)
4. Remind — Next conversation, agent reminds the user and asks for permission
5. Install — After approval, agent installs the skill and activates it

All created skills are pushed to the FutureHax GitHub organization following the openclaw-skill-<name> naming convention.

Installation

Copy the skill to the OpenClaw shared skills directory:

# Shared (all agents)
scp -r skill-manager your-vps:~/.openclaw/skills/skill-manager

Restart the gateway after installing:

openclaw gateway restart

Verify installation

openclaw skills list | grep skill-manager

Requirements

• GITHUB_TOKEN — GitHub PAT with read-only access (checks repo existence, clones public repos). This is the always-on token that keeps the skill in ready state.
• GITHUB_WRITE_TOKEN (optional) — GitHub PAT with repo scope for the FutureHax org. Only needed when create.sh runs to create repos and push. Can be short-lived or limited-scope. If not set, create.sh falls back to GITHUB_TOKEN.
• python3 — Used for JSON manipulation in tracking scripts (standard on Ubuntu 24.04)
• git — For initializing and pushing skill repos
• curl — For GitHub REST API calls

Add the tokens to ~/.openclaw/.env:

GITHUB_TOKEN=ghp_your_readonly_token
GITHUB_WRITE_TOKEN=ghp_your_write_token

Skill contents

skill-manager/
├── SKILL.md           # Skill definition and agent behavioral instructions
├── README.md          # This file
└── tools/
    ├── create.sh      # Create GitHub repo, push skill, track as pending
    ├── pending.sh     # List/check/add/remove pending skills
    ├── install.sh     # Install approved skill from GitHub
    └── status.sh      # Check skill existence (installed, GitHub, pending)

Tool usage

Check skill status

bash tools/status.sh weather-api

Returns:

{
  "name": "weather-api",
  "installed": false,
  "github": false,
  "pending": false,
  "paths": []
}

Create a new skill

Stage files first, then run create:

# Stage the skill files
mkdir -p ~/.openclaw/skill-manager/staging/weather-api/tools
# ... write SKILL.md and tool scripts to staging dir ...

# Create repo and push
bash tools/create.sh weather-api "Query weather data for location-based recommendations"

Returns:

{
  "status": "created",
  "name": "weather-api",
  "repo": "FutureHax/openclaw-skill-weather-api",
  "url": "https://github.com/FutureHax/openclaw-skill-weather-api",
  "pending": true
}

Check for pending skills

bash tools/pending.sh check

Returns nothing if no pending skills, or a formatted list:

- **weather-api**: Query weather data for location-based recommendations
  Repo: https://github.com/FutureHax/openclaw-skill-weather-api
  Created: 2026-02-12T20:00:00Z

Install an approved skill

# Shared (all agents)
bash tools/install.sh weather-api shared

# Single agent only
bash tools/install.sh weather-api agent:zordon

Returns:

{
  "status": "installed",
  "name": "weather-api",
  "path": "/home/marvin/.openclaw/skills/weather-api",
  "scope": "shared",
  "repo": "FutureHax/openclaw-skill-weather-api",
  "note": "Restart the gateway to activate: openclaw gateway restart"
}

How the agent uses this

The agent follows a strict workflow defined in SKILL.md:

1. Conversation start: Runs pending.sh check to see if any skills are awaiting approval
2. Capability gap: When the agent can't fulfill a request, it proposes creating a skill
3. Creation: Agent writes SKILL.md + tools to a staging directory, then runs create.sh
4. No auto-install: Created skills are never installed in the same conversation
5. Next conversation: Agent reminds the user about pending skills and asks permission
6. Approval: User approves, agent runs install.sh, suggests gateway restart

License

MIT