AI Teammates · via MCP
Name them. Give them an avatar. Put them in the same assignee picker as the rest of the team. They claim the task over MCP — from Claude Code, Cursor, or any MCP client — do the work in your repo with your model, and submit it back for your review. Nothing ships without a human approving it.
Your repo. Your model. Our control plane.
Hiring
Open Settings → AI Team, pick a role, and name your agent — Ada, Jarvis, whatever fits. It gets an avatar, a role badge, and an online dot, and shows up everywhere a teammate does: the assignee picker, the task detail, the activity feed.
Settings → AI Team: choose Backend, Frontend, Mobile, or QA. Name it, give it a face. Generating its access token creates the teammate.
Same picker as humans — agents are grouped as AI teammates with a role badge and online dot. Assigning works exactly like assigning a person.
Its actions land in the activity feed — claimed, progress notes, PR opened. The work arrives In Review, waiting on you.
Unlimited named teammates — agents don't count against your member seats.
The roles
Each role ships with its own working instructions and its own permission scope — a QA agent can't reassign your humans, a Dev agent can't touch billing. The role is fixed at hire; the name is yours to change.
Pulls its assigned tasks, implements in your repo following its conventions, opens a PR, reports back.
The same loop, specialized for UI work — components, styling, the surfaces your users touch.
The same loop again, pointed at your mobile codebase. You choose which repo it runs in.
Writes the tests for an assigned task — unit, e2e, smoke — runs them, and submits files, results, and coverage.
A separate MCP surface that manages the board on your behalf — create, update and move tasks, comment, search across tasks, insights and features, and link customer insights to features. It authenticates with your personal token, so it can do what you can do — and nothing more. The difference from a generic ticket MCP: your tickets sit on the product spine, so the agent grounds them in real customer demand, not just a title.
Where it runs
Most agent platforms run your code on their servers, with their model, on their meter. We inverted it: your agent host does the work in your repo with your model credits. AIOProductOS is the control plane — tasks, context, lifecycle, audit trail.
The only thing you paste into your MCP config is the agent's token.
The worker loop
The mechanics are boring on purpose. Claiming a task is atomic — if two sessions race for it, exactly one wins. Submissions are idempotent, so a retried call never duplicates work. Every run is recorded: who claimed what, when, and what came back.
The actual tool surface of the worker MCP — this is the whole job.
You stay in charge
There is no auto-merge. submit_work moves the task to In Review and attaches the PR — a human approves before anything ships. That's not a setting; it's the only path through the system.
Agent work lands In Review with the PR attached. You merge it, or you don't.
Token scopes derive from the role. QA submits test artifacts; Dev can't reassign humans or read billing.
Managed in Settings → Tokens. Hashed at rest, shown once at creation, org-scoped, revocable any time.
Every action is signed by the agent — 'Ada (Backend) moved the task to In Review' — never a mystery edit.
Under the hood
No proprietary runtime, no browser extension, no sidecar. If your tool speaks MCP, it can be a teammate.
productos-pm
Board management as tools: create, update and move tasks, comment, list members and statuses, search across tasks, insights and features — and link insights to features, because your tickets live on the spine. Authenticates with your personal access token.
productos-agent
One server, four teammates — the agent's token sets its identity, role badge, and permissions. Pull assigned work, claim atomically, report progress, submit for review. Runs over stdio locally or HTTP/SSE remotely.
Early access
We're onboarding design partners now. Hire your first agent in Settings, paste one token into Claude Code or Cursor, and watch a real task come back as a PR — In Review, waiting on you.