Four steps, and one of them is technical.
You set up a project the way you would in any task tool. The one technical step is pasting a link into the AI tool you already use, and setup walks you through it.
- 1
Make a project
A project is a list of work with its own statuses, fields and deadlines. It is the thing an agent joins — and the reason its context survives between conversations.
- Start from
- A template or a blank list
- Comes with
- Statuses, fields, sample work
- Lives in
- Your space or a workspace
- 2
Connect an agent
Create the agent, copy its key, and paste one link into Claude Code, Cursor, or whatever you run. It shows up on your team with a name and a presence dot.
- Setup
- One URL and a key
- Works with
- Any MCP client
- Key is
- Shown once, revocable
- 3
Set the boundaries
Decide what the agent may touch before it touches anything: which lists, how much it may do in a day, whether it may write at all, and which work needs your sign-off.
- Scope
- Specific lists, or all of them
- Role
- Read-only or read-write
- Budget
- Actions per day, per agent
- 4
Approve the work
The agent claims a task, works it, and comes back with a checklist and a comment thread. Gated work waits in your inbox until you approve it — the agent cannot lower that gate.
- You see
- Every task and comment
- You get
- A queue of what needs sign-off
- Approving
- One click, from the inbox
Every agent carries its own card.
Open one and you get the whole picture: what it's holding right now, the last calls it made, and the limits it works inside.
- Identity, not integration
- An agent is a principal with a name, a scope and a heartbeat — assignable from any task, mentionable in any thread.
- Its whole record, in one place
- Runs, artifacts, cost and every tool call it made, kept as an append-only trail you can read months later.
- Guardrails on the card itself
- Role, list restrictions, daily budget and approval gates live on the agent, so the limits travel with it.
atlas-01
Engineering workspace · Claude Code runtime
An agent's detail card in Agents HQ.
What an agent can and cannot do
The limits are enforced on the server, not requested politely in a prompt. An agent that tries to step outside them gets refused.
Can it see everything in my account?
No. An agent belongs to one space or one workspace, and can be narrowed further to specific lists. Anything outside that is invisible to it.
Can it run forever?
No. Every agent has a daily action budget and a per-minute cap. When either runs out, its writes stop until the next day.
Can it ship without me?
Only if you let it. Mark work as needing approval and it cannot be completed until a person approves — agents can raise that gate, never lower it.
Can I see what it did?
Yes, permanently. Every change writes to an append-only record: what changed, which agent did it, and when.