Features
Devices
Cloud machines are great until the thing you need is on your desk. The iOS simulator. A database you only run locally. A service behind your company VPN. A board plugged into a USB port. An agent in the cloud can't reach any of that.
That's what Devices are for. If the Sheepdog desktop app is open on your computer, an agent can run commands on it. By default it asks you first, every time.
Connect a computer
This takes about a minute. Install the desktop app and sign in. Then go to Settings → Devices and press Connect this computer. (There's also Agents → Let Agents Use This Computer… in the menu bar, if you like menus.)
It'll ask you to pick a working folder, which is where commands start. Your home folder is the default.
That's it. You don't open any ports or install a daemon. Your computer stays connected as long as the app is open. If you close it, agents are told the computer's offline and get on with things in the cloud.
Use it in a task
You usually don't have to do anything. When you have a device connected, agents know about it and use it when the work calls for it. If you want to be specific, @mention the device in your message.
Behind the scenes, the agent uses a command called device:
Terminal
device status
device run --name "iOS build" -- xcodebuild -scheme App build
Commands run in your login shell, from your working folder, as you. They see your files and your logins. That's the whole point, but it's also why the next section matters.
Saying yes to commands
Each command shows up as a card in the task's chat, with the exact command the agent wants to run. You get a notification too, so you can answer from your phone.
You get three choices. Run runs that one command. Always allow for this task runs it and lets the agent skip asking for the rest of this task. Don't run says no, and the agent hears about it.
Only you can answer for your own computer. If nobody answers in 10 minutes, the command just doesn't run.
Actually read the command before you say yes. Agents read web pages, files and tool output, and any of those can try to talk an agent into running something you never asked for.
Long-running commands
Commands don't have to be quick. A dev server, a file watcher or a 40-minute build all keep running until they finish or someone stops them. The agent can come back later to check on them, and it's told to clean up after itself.
You can watch, too. Once a task has run something on your computer, it gets a Device tab with each command's live output. It works from any browser or your phone, and it has a Stop button. The desktop app's Agents menu shows what's running as well.
Full access
If saying yes all the time gets tedious, you can set a computer to Full access in Settings → Devices. Commands from your own tasks then run straight away, with no card in the chat. The computer still shows a notification for each one, so nothing happens silently.
Automations need this, since nobody's around at 3am to say yes.
The honest advice: only use full access on a machine you've set aside for agents. Not the laptop with your bank logins on it.
Sharing with your team
Say your team has a Mac mini in the corner with every simulator installed. You don't want everyone setting up their own.
Share it from Settings → Devices, and anyone in your organisation can @mention it in their messages. Commands run as you. You choose whether you approve theirs:
- Team · ask me: each command waits for your yes. If you can't see the task they're asking from, the card lets whoever started it add you in one click.
- Team · full access: their commands run straight away, with no card in the chat. Only for a machine set aside for the team.
This is separate from Full access, which only covers your own tasks. So you can skip the questions yourself and still approve everyone else's.
Disconnect a computer
Press Remove next to it in Settings → Devices, or pick Agents → Disconnect This Computer… in the desktop app. Anything still running on it stops.