Your first task

For your first coding task, you'll ask an agent to make a small change: you'll give the task its own devbox and thread, describe the change, check the result yourself, and save the code to your Git remote.

Start after project setup is complete and your app runs in the project's saved environment, because that environment is what a new devbox will start from. If you created the project from GitHub, cloning the repository was only the first step; set up the development environment first so a devbox can run your app.

A small task to start with

This guide follows one example task: a web app with a search bug. When you clear the search field, the old results stay on the page; they should disappear. The problem is easy to describe, the change is small, and you can check the fix yourself by using the app.

You can follow these steps with a different task. Choose a small change that you can describe clearly and check yourself, and substitute your task wherever this guide uses the example.

Start the conversation

1. Select your project in the desktop app.
2. Choose New devbox + thread, then select Codex or Claude. If the reply box shows Set up Codex or Set up Claude, use that button to connect the agent first.
3. Describe the problem, the change you want, and how the agent should test it, then send your request.

For the example bug, a useful first prompt looks like this:

When I clear the search field, the results stay visible. Fix it so clearing the field also clears the results. Run the relevant tests, check the behavior in the app, and explain how I can review the change.

Describe the search bug, the expected behavior, and the checks you want the agent to run. Select the image to enlarge it.

boxes.dev will prepare a devbox from your project's saved environment and start the agent there, working in that devbox's checkout of your code. Read its progress in the thread, answer its questions, and send corrections or follow-up instructions in the same conversation.

Check the result

Keep the agent's thread selected so the titlebar tools open its devbox; the desktop tour shows where each control lives.

1. Open Git/Review in the titlebar and read through the changed files. Ask the agent to explain anything unclear and to report which checks passed.
2. Ask the agent to start your app if it isn't already running. The new devbox runs its own processes, and its startup script may already have started the app.
3. If the change shows in your app's pages, open Ports and select the web address of your app's development server; it will open in your laptop's browser. If it doesn't — a command-line fix, say — ask the agent to run the command or the relevant test, and read the output yourself.

In the example, check all three parts in the browser: a search shows results, clearing the field removes them, and a fresh search still works.

If no port appears, ask the agent to check that the server started on this devbox. For a port warning, open the Ports menu to see the cause and the available action; see Preview your app.

If the change needs more work, describe what happened and what you expected in the same thread, including the page or file involved, then review the revised result.

Open the app's address from Ports to check the result in your browser. Select the image to enlarge it.

Save your work

The agent's changes are ordinary edits in the devbox's checkout. Use Git/Review to commit and push them, and open a pull request if your team uses one; Commit, push, and open a pull request walks through the steps. Changes on this devbox won't automatically appear on your Template box or other devboxes, so pushing is how the work leaves the machine.

To return to the task later, open the devbox's status menu and choose Sleep devbox. Your files and conversation will be there when you wake it. When the task is finished and pushed, see Manage devboxes to delete a devbox you no longer need.