Agents don't open your dashboard. They run commands, read the output, and retry when something fails. Without a CLI, they guess at endpoints, auth, and pagination, and fail quietly.
Docs and OpenAPI specs describe the shape. They don't give an agent a command it can run, retry, or verify.
Agents speak shell
A command line is the one interface every coding agent can already drive. No SDK, no docs crawl, no browser.
Mapped before written
Reads, writes, auth, pagination, and errors are planned in CLI.md before code, so nothing is improvised at runtime.
Proven before shipped
No TTY prompts, stable --json, real exit codes, and dry runs on writes. The conformance checker proves it before an agent trusts it.
Every request is a command.
The skill installs once and stays in your agent's skill folder. After that, every request is a command, not a conversation.