Teach your agent
new skills.
Skills are scoped system prompts that give your OpenClaw agent domain-specific knowledge. Install from ClawHub or write your own.
What is a skill?
A skill is a SKILL.md file that contains a scoped system prompt. When activated, it augments your agent with specialized instructions for a specific domain.
Skills can define tool preferences, output formats, personality traits, or domain expertise. They compose with each other and the base agent personality.
Think of them like plugins for your agent's brain — each one teaches it how to handle a specific type of request.
# SKILL.md — Recipe Assistant
You are a cooking assistant. When the user
asks about recipes:
- Suggest ingredients based on dietary
preferences stored in vault
- Provide step-by-step instructions
- Include prep time and serving size
- Convert measurements on request
## Tone
Friendly and encouraging. Use simple
language. Avoid jargon.Built-in skills
Every OmniClaw agent ships with these skills pre-installed and active.
General AssistantBuilt-in
Core conversational ability. Handles greetings, general questions, small talk, and routing to specialized skills.
Web SearchBuilt-in
Searches the web when the user asks about current events, facts, or anything requiring up-to-date information.
Code RunnerBuilt-in
Writes, executes, and debugs code inside the sandbox. Supports Python, Node.js, and shell scripts.
File ManagerBuilt-in
Reads, writes, and organizes files in the sandbox filesystem. Handles uploads and downloads via WhatsApp.
Install community skills.
ClawHub is the community registry for OpenClaw skills. Browse verified skills, check ratings, and install with a single command.
$ openclaw skills install translator
Installing translator@1.2.0...
✓ Skill installed
$ openclaw skills list
general-assistant (built-in)
web-search (built-in)
code-runner (built-in)
file-manager (built-in)
translator 1.2.0Skills are sandboxed inside your agent's VM. They can access vault secrets (with your permission) and the local filesystem, but never other users' data.
$ openclaw skills install calendar
Installing calendar@0.8.1...
⚠ This skill requests vault access:
- GOOGLE_CALENDAR_KEY
Allow? [y/N] y
✓ Skill installedCreate your own skills.
Creating a skill is as simple as writing a markdown file. Drop a SKILL.md in your agent's workspace directory and it's available immediately.
OmniClaw hot-reloads skills when files change — no restart needed. Edit the file, send a message, and your agent uses the updated instructions.
Skills live in /.openclaw/skills/ inside the sandbox. Each skill gets its own directory with a SKILL.md and optional configuration.
# Directory structure
/.openclaw/skills/
├── general-assistant/
│ └── SKILL.md
├── web-search/
│ └── SKILL.md
├── my-custom-skill/
│ ├── SKILL.md
│ └── config.json
└── ...
# config.json (optional)
{
"name": "my-custom-skill",
"version": "1.0.0",
"vault_access": ["MY_API_KEY"],
"triggers": ["when user asks about X"]
}Skill supply chain matters.
The risk. Research shows roughly 12% of community-published AI agent extensions contain prompt injection or data exfiltration patterns. Skills have broad access to your conversations and can influence what your agent says and does.
OmniClaw's approach
- Skills run inside your isolated microVM — they cannot access other users or the host system
- Vault access requires explicit user approval per-secret
- Agent sandboxes have internet access with open outbound egress, so review what a skill does before you install it
- ClawHub verified skills are manually reviewed before publishing
- VMs auto-destroy after 24h — even if a skill is malicious, the damage is contained and temporary
Build your perfect agent.
Start with built-in skills, install from ClawHub, or write your own. Your agent, your rules.