Press ⌘ J anywhere.
What it can actually do
It reads and writes your workspace: create boards and cards, move and complete them, plan a week, write and edit notes, log a habit, record an expense, add elements to a whiteboard, search across everything, and answer questions using the Mind.
It does this through exactly the tool surface an external agent gets. There is no privileged internal path, which is why the assistant cannot do anything you could not do yourself through the interface, and why anything you can do is also available to an agent you connect. See Agents and MCP.
Specialist agents
Alongside the general assistant there are agents with a narrower job: a planner for laying out a week, an inbox agent for triaging captures, a research agent that can search the web. You can invoke one directly or let the assistant hand work to it.
Workspaces on paid tiers can define custom agents with their own instructions and tool access, for the workflow you run constantly and describe identically every time.
What an action costs
Every tier includes a monthly allowance of AI actions: 40 on free, 400 on Pro, 1,200 pooled across the household on Family.
One action is roughly one thing you asked for. A question the assistant answers by reading is one action; a request that has it do a piece of work involving several steps costs more, because a turn that calls tools is genuinely a loop rather than a single call.
The current usage is shown in Settings. When the allowance runs out, the assistant stops rather than silently degrading. An assistant that quietly gets worse near the end of the month is harder to trust than one that says it has stopped.
What it is not allowed to see
The Vault, while locked, and anything marked Hide from memory. Not filtered out of the answer at the end: never retrieved in the first place. See The Mind.
In a shared workspace, the assistant is scoped to what you can see. It cannot be used to read a household member's private material.
Where it runs
The assistant runs as a separate service, not inside the web app. This is why the rest of the product stays fast while a long task is in flight, and why the model behind it can be replaced without touching the app.
It also means the assistant needs a network connection. Everything else in the app works offline; this does not.
When it gets something wrong
It will. The useful correction is usually to the Mind rather than to the conversation: if the assistant keeps believing something untrue about a project, the fix is to open the Mind page and delete the memory saying so. The conversation resets; the memory does not.