Skills, Tools & Plugins

Teach your agents new tricks. Write a skill in Markdown, add a tool in JavaScript, or install a whole package from GitHub with one command.

Four kinds of extensions

Where to put them

Each type loads from three tiers. When names clash, workspace beats global, which beats built-in.

After editing, run /skills reload, /tools reload or /plugins reload. You don't need to restart.

Writing a skill

# Skill: release_notes
## Description
Write customer-facing release notes from recent commits.
## Trigger
When asked for release notes or a changelog.
## Instructions
1. Run git_log for commits since the last tag.
2. Group changes into Added / Changed / Fixed.
3. Write in plain language for non-technical readers.
## Output Format
Markdown, saved to RELEASE_NOTES.md.

Writing a tool

export const description = {
  name: 'my_tool',
  description: 'What this tool does',
  parameters: {
    type: 'object',
    properties: { input: { type: 'string', description: 'The input value' } },
    required: ['input'],
  },
};

export async function execute({ input }) {
  return `Result: ${input}`;
}

Writing a plugin (slash command)

export const description = {
  command: 'standup',
  usage: '/standup',
  description: 'Summarize yesterday\'s work for standup',
};

export async function execute(args, { emit, runIteration }) {
  emit('info', 'Preparing standup…');
  await runIteration('Summarize what changed in this workspace yesterday as standup notes.');
}

Plugins receive workspaceDir, agentosDir, emit, sendMessage and runIteration in their context.

Installing from GitHub: /install

Point AgentOS at a repository and it works out what to install. That can be skills (SKILL.md), Claude Code plugins, MCP servers, CLIs or AgentOS tools and plugins. It shows you a plan first, and nothing is installed until you confirm.

/install anthropics/skills --only pdf,docx
/install confirm

Options: --yes, --dry-run, --only / --skip, --env KEY=value, --global (install for every workspace) and --force. If a repo has no recognizable layout, the agent reads its README, proposes install steps and asks before running them.