Quick Index

• Why Cowboy
• What Makes Cowboy Different
• Features
• Installation
• Quick Start
• Core Workflows
• How Cowboy Works
• Documentation
• License

Why Cowboy

AI coding agents get dramatically better when they have strong, task-specific instructions.

But today, skills are:

• copied manually between repos
• duplicated across agent ecosystems
• outdated after a few weeks
• hard to maintain
• impossible to version properly

Cowboy fixes this.

It gives you a single user-level source of truth for skills that can be:

• installed from GitHub
• generated automatically from docs or repos
• added to multiple projects
• synced to multiple agents
• updated to latest versions
• enabled or disabled per agent
• maintained over time

All with a simple CLI.

And most importantly:

Cowboy uses your existing Codex or Claude Code subscription. No API keys. No extra cost. No vendor lock-in.

What Makes Cowboy Different

Auto-updating skills

Install a skill once. Update it anytime.

cowboy update

Cowboy re-fetches imported skills and regenerates AI-authored ones using the latest docs or repos.

Your agents always stay up to date.

Generate skills automatically

Create new skills from:

• GitHub repositories
• documentation URLs
• local documentation directories
• free-text topics

cowboy generate playwright testing
cowboy generate --repo https://github.com/langchain-ai/deepagents
cowboy generate --docs https://docs.stripe.com/api
cowboy generate --docs ~/docs/stripe-api
cowboy generate --repo https://github.com/langchain-ai/deepagents --docs ~/docs/deepagents

Cowboy uses your local agent CLI to generate the skill — completely free.

Multi-agent sync

Install once. Use everywhere.

Cowboy syncs the same skill to:

• Claude Code
• Codex

From one canonical package.

One canonical source of truth

Skills live once in the user's global Cowboy data directory:

• macOS: ~/Library/Application Support/Cowboy
• Linux: $XDG_DATA_HOME/cowboy or ~/.local/share/cowboy
• Windows: %LOCALAPPDATA%\\Cowboy

Projects keep links in .cowboy/skills/{name}/, and agent directories also use links instead of copies.

No drift. No duplication. No manual copy-paste.

Features

| Capability | What Cowboy does | | ------------------------ | -------------------------------------------------- | | Auto updates | Refresh skills from GitHub or regenerate from docs | | AI skill generation | Generate skills from repos, docs, or topics | | Multi-agent install | Sync skills to Claude Code and Codex | | Canonical storage | One source of truth in the global Cowboy library | | Repo import | Install skills from GitHub skill repositories | | Free generation | Uses your own Codex / Claude Code subscription | | Explicit install targets | Generate with one agent, install to another | | Enable / disable | Toggle skills per agent without deleting | | Version control friendly | Projects keep lightweight links instead of copies | | No vendor lock-in | Works with local agent CLIs |

Installation

npm install -g cowboy-cli

For a project-local install:

npm install cowboy-cli
npx cowboy

Requirements

• Node.js >= 20.12.0

• At least one supported agent CLI:

  • Claude Code
  • Codex

The published package name is cowboy-cli, and the executable is cowboy.

Quick Start

cowboy init
cowboy install https://github.com/ComposioHQ/awesome-claude-skills
cowboy generate deepagents
cowboy list
cowboy list --all
cowboy update

Core Workflows

Install skills from GitHub

cowboy install https://github.com/ComposioHQ/awesome-claude-skills
cowboy install <repo> --install-for claude --install-for codex

Generate skills automatically

cowboy generate langchain
cowboy generate --repo https://github.com/langchain-ai/deepagents
cowboy generate --docs https://playwright.dev/docs/intro
cowboy generate --docs ~/docs

--docs now accepts either a documentation URL or a local directory path. Local directories are exposed directly to the agent session so it can build the skill from files on disk as well as web docs.

Cross-agent example:

cowboy generate --repo https://github.com/stripe/stripe-node --agent codex --install-for claude

Runtime override:

cowboy generate \
  --repo https://github.com/langchain-ai/langgraph \
  --agent codex \
  --effort xhigh

Update skills

cowboy update
cowboy update playwright-testing
cowboy update --agent codex --effort high

How Cowboy Works

Cowboy manages skills in three layers:

1. Discovery or generation

• Imported from GitHub
• Generated from docs/repos/topics

2. Canonical storage

<global Cowboy data dir>/skills/{name}/

This is the real canonical package.

Projects keep a local alias at:

.cowboy/skills/{name}/

3. Agent sync

.claude/skills/
.agents/skills/

These project-local agent directories are links to global agent views, not copied folders.

Skill Package Format

skill-name/
├── SKILL.md
├── scripts/
├── references/
└── assets/

Example:

---
name: playwright-testing
description: Reliable browser testing workflows with Playwright
---

# Playwright Testing

Prefer deterministic selectors and isolate flaky test causes before retrying.

Project Layout

your-project/
├── .cowboy/
│   ├── config.yaml
│   ├── installed.yaml
│   └── skills/        # links to the global library
├── .claude/
│   └── skills/        # links to global Claude views
└── .agents/
    └── skills/        # links to global Codex views

Supported Agents

| Agent | Install location | | ----------- | ------------------------ | | Claude Code | .claude/skills/{name}/ | | Codex | .agents/skills/{name}/ |

Cowboy stores canonical skills globally and links them into each project.

Commands

cowboy init
cowboy install <github-url>
cowboy add <name>
cowboy generate [topic]
cowboy update
cowboy list
cowboy list --all
cowboy enable <name>
cowboy disable <name>
cowboy remove <name>
cowboy agents
cowboy default-agent <agent>

Documentation

See docs/:

• docs/index.md
• docs/commands.md
• docs/generation.md
• docs/architecture.md
• docs/maintainers.md

Open Source

Cowboy is structured for external contributors and maintainers:

• CONTRIBUTING.md for the contribution workflow
• CODE_OF_CONDUCT.md for community expectations
• SECURITY.md for responsible disclosure
• CHANGELOG.md for release history
• docs/maintainers.md for branch protection and release operations

Contributing

See CONTRIBUTING.md before opening a pull request.

High-signal defaults:

• Keep skills portable
• Avoid agent-specific assumptions in core flows
• Add tests for behavior changes
• Update docs when CLI behavior changes

pnpm test
pnpm build

License

MIT