Run an agent on another machine from your browser

The big machine is under your desk and you are on a laptop at the kitchen table. Start the agent on the workstation from your browser, answer the one question it has, and close the lid. Any coding agent the workstation has installed can be the one you start; the example uses Claude Code

Pick a machine and a task in the browser, answer the agent's one question, close the lid, come back to an open pull request. Everything below stands on its own

Where else this applies

You are travelling. The workstation at home keeps a long task moving while you check in from a hotel, a train or a phone
An urgent fix from the soccer field. Open the browser, start the agent on the machine that has the repository and the credentials, answer its one question, and follow the pull request from the sideline
Your main machine is in the repair shop. A colleague’s daemon, or a spare box with the repository cloned, carries on while yours is away
The build needs the big machine. A laptop that cannot run the test suite or the model locally can still drive the machine that can
You want to run two things at once. Start a second agent on another machine instead of waiting for the first to finish on this one

1. Once: make the workstation available

On the workstation, with kcap set up and the coding agent you want installed and signed in:

kcap daemon service install
kcap daemon service status

The service starts at login and restarts after a crash. If you used the desktop app’s setup, it offered to do this for you. Open Agents in the web app from any machine: the workstation is listed with the vendors it can launch

2. Launch from the laptop

In Agents, choose New agent. Pick the workstation as the daemon, the vendor, a repository from the ones the daemon knows, a model if you want one, and write the prompt:

Investigate why checkout retries sometimes create duplicate orders. Explain the cause, implement a fix, and run the relevant tests

Set Sharing if a teammate should be able to follow along, then launch. The agent starts in a fresh git worktree on the workstation; your checkout there is untouched

3. Answer the permission request

When the agent needs permission, the request appears in its card and as a notification in the web UI. Answer Yes, Yes, always or No. A request waits for you; it does not time out while you are at lunch

4. Come back later

The conversation is where you left it, whether you reopen it on the laptop, on the workstation, or in the desktop app. Send another message, ask for a summary, or Stop it. The session is stored like any other, so the branch and its reasoning are in Sessions afterwards