Bring your own AI agent

Waymarq's in-app AI reads a spec and writes the breakdown for you. This is the same idea from the other direction: your own agent does the reasoning, and Waymarq is just the structured place it writes fully-formed data into. That means it works even when an organisation has no AI provider configured in Waymarq at all — the intelligence is entirely on your side of the connection.

The pattern that works well, in order:

1. Fetch /api-docs first — that's the reference it should work from, not this page.
2. Create the project, then phases, then tasks under each phase — carrying a real markdown_spec and llm_test_plan per task, not just a title.
3. Attach a test_plan checklist to anything worth verifying, and split a task into subtasks only when the work is genuinely independently trackable — not by default.
4. Call POST /projects/{project}/schedule once every task has an estimate, so the project has real dates rather than none at all.
5. As work happens, keep it current — status through the dedicated endpoint, comments for context, time logged as it goes.
6. Write the documentation as features land. GET /tasks/{task}/documents shows what already covers an area before you change it, and GET /projects/{project}/documents/coverage shows what nothing describes yet — so the person you're working for stops having to ask how their own software behaves.

There's a prompt generator under Organisation settings → API → Import guide that builds a ready-to-paste brief for your agent, pointed at your account and — optionally — a specific existing project. It bakes in the two rules that actually matter (fetch the docs first, never write status directly) so you don't have to remember to say them yourself.

If your agent would rather run a command than make HTTP calls, there's a CLI — npx @waymarq/cli next, start, done, comment, log. Every command takes --json. It works anywhere there's a shell, which is everywhere: Claude Code, Codex, Cursor, a CI step, a git hook, or you.

Commit the .waymarq.json it gives you into your repository root and the binding stops depending on that prompt. It names the project, nothing else — no credentials, since it's a committed file — so any session, any machine and any teammate's agent picks up where the last one left off instead of having to be told again.