@package-system

Commands

commands[] — prompt commands a package serves as static MCP prompts to external clients.

A command is a ready-made prompt an external MCP client can list and fetch. Each commands/*.json file is one entry, and it must carry an id; a file without one, or one that does not parse, is skipped with a warning. Inline commands[] entries in the manifest merge with the files, and a file wins on the same id.

{
  "id": "weekly-summary",
  "name": "weekly-summary",
  "kind": "prompt",
  "description": "Summarise this week's project activity.",
  "messages": [
    { "role": "user", "content": { "type": "text", "text": "Summarise this week's activity in the project." } }
  ],
  "requires": { "features": ["acme.reports.read"] }
}
  • Only kind: "prompt" commands are served, as MCP prompts (prompts/list, prompts/get). What a client receives is the description and the messages array, verbatim.
  • There is no executor: declared arguments are not substituted into the messages, and a kind: "script" command is never served. Anything that must run belongs in a skill or a tool.
  • requires.features gates the command the same way it gates every other element: a caller who lacks the feature never sees it in prompts/list.
  • The package's runtime.hosted.prompts setting (all, selected or none, see the manifest) decides whether its commands are hosted at all.

A markdown file directly inside commands/ is a different thing: the older slash-command form, loaded as a skill whose id is the file name.