Skills System

Skills are the mechanism by which AnyCowork agents gain new capabilities. While "Tools" are low-level primitives (like reading a file or searching text), "Skills" are higher-level bundles of functionality that solve specific user problems.

What is a Skill?

A Skill in AnyCowork is a package containing:

1. Metadata: Name, description, triggers, and configuration.
2. Implementation: Scripts (Python, Node.js, Shell) or native commands.
3. Documentation: Instructions for the AI on how and when to use the skill.
4. Security Config: Sandbox requirements and permission scopes.

Examples:

• PDF Generator: A skill that takes markdown text and converts it to PDF using a Python script.
• Data Analyst: A skill that uses Python's Pandas library to analyze CSV files.
• Git Ops: A skill to check status, commit, and push code.

Sandboxing & Security

One of the unique features of AnyCowork is its Docker-based Sandbox. This allows agents to execute untrusted code (like Python scripts generated on the fly or downloaded skills) safely without risking your host machine.

Execution Modes

Skills and Agents can operate in three execution modes:

1. Sandbox (Recommended for most skills)

   • Code runs inside an isolated Docker container (debian:stable-slim or alpine).
   • The container has limited resources (RAM/CPU) and restricted network access (optional).
   • Your workspace is mounted, but system files are inaccessible.
   • Requires: Docker installed and running on the host.

2. Flexible (Default)

   • If Docker is available, it uses the Sandbox.
   • If Docker is missing, it falls back to running on the host, but may warn or fail if the skill requires sandbox.

3. Direct

   • Runs directly on your host machine.
   • Useful for skills that need host access (e.g., managing system services, accessing hardware).
   • Security Warning: Only enable this for trusted skills.

Enforcing Safety

You can configure an Agent to force Sandbox mode via its Execution Settings. In this mode, if Docker is unavailable, the agent will refuse to execute skills rather than falling back to unsafe direct execution.

Developing Skills

Skills are stored in the skills/ directory (either bundled with the app or in your user data folder). Each skill is defined by a SKILL.md file with YAML frontmatter:

my-skill/
├── SKILL.md              # Definition (frontmatter + instructions)
└── src/
    ├── main.py           # Implementation
    └── requirements.txt  # Dependencies

SKILL.md Format

Skills use a single Markdown file with --- YAML frontmatter for metadata:

---
name: pdf-generator
description: Converts Markdown to PDF using python-markdown and weasyprint.
license: Apache 2.0
---
 
# PDF Generator Skill
 
Instructions for the AI on how and when to use this skill...

The body of the Markdown file serves as both documentation and AI context - the agent reads it to understand how to invoke the skill.

For a complete guide on adding skills, see Extending AnyCowork.