agents/: pilots for agents

Everything an agent needs to use pilots, in one directory that is three packages at once:

• A Claude Code plugin. .claude-plugin/plugin.json, .mcp.json (the hosted MCP server), hooks/hooks.json with scripts/guard.py (asks before destroy_machine, restore, rollback, and before an exec that should be checkpointed first), and the skills under skills/: pilots (loaded by the model when the task is about pilots), status (/pilots:status) and smoke (/pilots:smoke).
• A portable Agent Plugins v1 package. plugin.json, mcp.json and the same skills/, for Cursor, Codex, GitHub Copilot, Kiro and every other client that loads that format.
• A Go module (github.com/pilotsrun/pilots/agents) that embeds the skill and holds the MCP toolset in mcp/. hostd serves the toolset at /mcp on every host and pilot mcp serves it on stdio; both read the pages from this embed, so the plugin, the CLI and the fleet ship the same words.

Install

Claude Code:

/plugin marketplace add pilotsrun/pilots
/plugin install pilots@pilots

Then set PILOT_API_KEY (and PILOT_API_URL for a self-hosted fleet) in the environment Claude Code starts from, or run pilot mcp install claude-code, which writes the server entry with the key pilot login stored. /pilots:status checks the result.

Any other MCP client: point it at https://api.pilotrun.app/mcp with Authorization: Bearer <key>, or run pilot mcp install <harness>; pilot mcp install --list names the harnesses it knows.

The skill

skills/pilots/SKILL.md and its references/ are the one copy of the agent docs. pilot init copies them into a repository's .agents/skills/pilots/, pilot skill install links them into ~/.claude/skills/pilots, and the MCP servers expose them as pilots-docs:// resources and through the docs tool. Edit them here; nothing else holds a copy.