Getting started

1. Create an account

Open the app and sign up with just an email and an 8+ character password. Sessions are cookie-based and last 30 days. Your tasks are private to your account — there's no shared workspace unless you explicitly add one.

2. Quick-capture a task

Type naturally into the capture bar. Due dates, priority, project, estimate, and a agent assignment all parse out of one line:

renew insurance next friday at 9am, p1, ~30m #admin @errands

becomes a task titled “renew insurance”, due next Friday at 09:00, priority urgent, a 30-minute estimate, filed under the Admin project and tagged Errands.

A few of the patterns it understands:

• today, tomorrow, next friday, in 3 days, or an ISO date — sets the due date. Prefix it with by for a hard deadline, or when: / due: to pin a date the title would otherwise swallow
• at 3pm, 9:30am, at 17:30 — sets a time of day (a time on its own means today)
• p1 / p2 / p3, or urgent / asap / low priority / someday — sets priority
• @agent (also @claude, @chatgpt) or “assign to an agent” — assigns the task to an AI agent
• ~30m, ~2h, est: 45m — sets a length estimate
• #project-name — files it into a matching project
• @tag-name — adds each matching tag; a task can take several

Typing # or @ in the capture bar opens your projects and tags to pick from, and the panel underneath shows the fields the line will set before you press Enter. A# or @ that matches nothing is left in the title rather than creating a project or tag.

Anything left over becomes the task title. You can also use the mic icon to capture by voice.

3. Organize with projects and dependencies

Projects live in the left sidebar; click one to filter the list. Tasks can also nest under a parent (subtasks), and a task can be marked “blocked by” another — Done cycle-checks the chain and keeps blocked tasks out of the agent's ready queue automatically.

4. Hand work off to an AI agent

Set a task's assignee to an AI agent, generate an MCP token from the Account panel, and connect an agent. See MCP setup for the full walkthrough.

Say what “done” means before you hand it over. Write the task's acceptance criteria as a checklist and the agent ticks each condition as it meets them — so progress is a number you glance at rather than a comment thread you read.

5. Answer, then approve

Two things come back to you, and both gather in the Control Room.

• Questions. On a real fork — spending money, anything outward-facing, anything irreversible — the agent parks the question and moves on to the next task instead of guessing or stalling. It comes with the agent's own recommendation and reasoning, so answering is usually one click, and you can answer from a Slack DM. Answering hands the task back; it never closes anything.
• Finished work. A task without agent_ok can't be closed by an agent at all — it proposes completion and waits. Approve to finish, or send it back with feedback.

Inside a task, the agent's write-ups sit in the thread and the bookkeeping — claims, status changes, run traces — collapses into a History fold, so the thread stays readable.

6. Let it run overnight

Point a scheduled MCP client at Done and it works the queue unattended. In Account → Notifications you can cap how far one night goes — tasks, minutes, tool calls, and nights per week — and tell it to stay quiet while it runs, so questions collect on the board for one morning pass instead of arriving one ping at a time. See MCP setup for scheduling.