A Project runs on its own computer in the cloud, so closing your laptop doesn’t stop it. Running in the cloud also lets a Project run more agents in parallel than your laptop could support.
When something needs testing on your machine, the coordinator starts a local agent to run it there.
But, you need to be running the CLI locally with agent worker start (aka My Machine) for this to work.
Thanks for the question. You’re close: the local agent does run on a self-hosted worker on your computer (the same kind of worker that shows up under My Machines), but the CLI isn’t the only way to start one.
You can get a worker running in either of two ways:
From the Agents Window: open Settings → Agents and turn on “Enable local self-hosted workers” under Self-Hosted Workers (on some versions this toggle is called “Remote Control”). Cursor then starts a worker for each local repo folder you have open in the Agents Window, so make sure the Project’s repo is open there. No CLI needed.
From the CLI: run agent worker start in your repo checkout, as described on the My Machines page. This is handy for a devbox or any machine where the app isn’t open.
A few things to know:
Either way, self-hosted machines need to be allowed for your team in the Cloud Agents section of the cursor.com dashboard (“Enable Self-hosted Machines” or “Enable Remote Control for Team”).
The first time the coordinator wants to run something on a machine, it asks you to approve it.
The local agent works in that machine’s existing checkout and branch, and the computer needs to stay awake with Cursor (or the CLI worker) running while it works.
Let me know if the coordinator doesn’t find your machine after that, and I’ll take a closer look.
Thanks for the reply. It’s not recommended to do both 1 and 2 on the same computer, I would imagine? You could have contention between two agents over the same files?
Good instinct. I’d pick one or the other per checkout.
If you turn on “Enable local self-hosted workers” and also run agent worker start in the same repo folder, you end up with two separate workers rather than one. They show up as two entries under My Machines (the desktop one is named after the folder and your machine, the CLI one after your hostname), and the coordinator can send work to either. Nothing coordinates file edits between them, so two agents could end up editing the same files at the same time.
Worth knowing too: even a single worker on your own machine can take more than one agent at once, so the same applies if the coordinator starts several local agents on one computer.
What I’d suggest:
On a computer where Cursor is usually open, use the desktop toggle and skip the CLI there.
Use agent worker start on machines where the app isn’t running, like a devbox.
If you want agents working locally in parallel, give each one its own checkout, for example a separate git worktree per task, each with its own worker. With the desktop app, opening each worktree folder in the Agents Window starts a worker for it. With the CLI, each extra worker on the same computer also needs its own --data-dir, since a second worker won’t start on the default one.
You can also try asking the coordinator in your Project to run local work one agent at a time.