When does a coding agent help with an app you already have?
A coding agent can read the files that make your app work, edit them and run checks. Try one when you can name the behavior you want to change and have somewhere to test it. If your booking form says “Something went wrong” even when a time has simply been taken, that’s a good first job: show a useful message and keep the rest of the form working.
Should you use the app builder you have or a separate coding agent?
Try your existing builder first if it can make the change and let you test a copy. A separate agent helps when you need to follow how a change works across files or work on code outside the original builder.
Before switching, find out whether you can open and run the project without the old editor. An exported folder may still need sign-in, a database or private connection settings. If you can’t run a test copy yet, get that working before you let a new tool edit anything.
What do you need before the first edit?
Start from a version that works. Ask the agent to show where the error comes from and explain how you can see it yourself before you approve a fix. If you can’t reproduce it, you won’t know whether it’s gone.
- Open just the booking projectGive the agent the folder for this app. Keep a copy of the working version or record it in Git, a tool that tracks your file changes.
- Check the test connectionsBefore opening the test app, check that it uses made-up bookings and cannot send real reminders or take payments. If you can’t verify that, let the agent inspect the files without editing or running the app.
- Reproduce the problemIn the test app, pick a slot you already booked. Write down what the app shows and whether it saves anything.
- Agree on one visible changeAsk the agent to explain how it would show “Choose another time” while keeping the name the visitor typed. Keep the rest of the form unchanged.
Compare tools
Scroll the table sideways to read every column.
| Tool | Where you open itThe first place you open the files and start the work. | Where work runsWhether the agent uses your computer or a separate one. | What needs reviewThe edits and actions to inspect before accepting the work. | How usage is paid forAn allowance, metered usage, or another service account. |
|---|---|---|---|---|
| Claude Code | Open a local folder in Claude Desktop’s Code tab, a supported editor or the terminal. The desktop Code tab needs a paid plan. | Local sessions work with your folder. Cloud work runs separately and needs its own connections. | Check the active permission mode. In Desktop, Manual can ask about edits and commands; Auto can allow actions without your prompt. Checkpoints cover tracked file edits. | Use the subscription allowance or a separately billed Claude API account. Check which account is paying before a long task. |
| Cursor | Open the project in Cursor’s editor. Read the files and review the changed lines alongside the chat. | The editor works with files on your computer. Eligible Cloud Agents work on a separate project copy. | Ordinary edits can save immediately. Check the diff and command settings; a checkpoint can help undo Agent edits to files. | New paid plans include a shared weekly pool for supported models. Check eligibility and extra charges for third-party models. |
| Cline | Install Cline in a supported editor, or use its desktop or terminal option. Choose how it connects to an AI service. | Local apps work on your files. Cloud Agents use a selected GitHub repository. | Use Plan to agree on work before Act. Check Auto Approve and compare file changes before restoring anything. | Depending on the app, use Cline credits, ClinePass, a separate AI account or a model on your computer. Cloud choices are narrower. |
| Roo Code | Retired. Use the original notice if an old guide still tells you to install it. | The former VS Code extension, Cloud and Router are historical products. | Keep the project files, then check the permissions and recovery tools of whatever you choose next. | No current signup. The retirement notice explains how Roo Code planned to handle unused balances. |
How do you know the agent fixed the right problem?
Try a taken slot and an open slot. The first should explain how to choose another time and save nothing. The second should still save exactly one booking and show the confirmation.
Look at the changed files as well. A diff is a view of the lines added and removed. You don’t have to understand every line to ask why changing one message involved rewriting sign-in. Ask the agent to name the changed behavior, show where it changed and tell you which checks actually ran.
If a check couldn’t run, keep that uncertainty visible. A passing check can verify what it covers. It can’t substitute for trying the booking flow if the check never opened the form.
- Find the first error before changing more codeUse the error index to narrow down whether a failing check comes from the browser, sign-in or saved records.
What should an agent be allowed to touch?
For the booking change, the agent needs the app’s files and somewhere safe to run them. It doesn’t need the account that sends real reminders, a list of real customers or access to every project on your computer.
Review the active approval settings before the task, especially what it can read, edit or run without asking. If it asks to run a command you don’t understand, ask what it will change and which account or database it will use. Stop if the answer doesn’t make sense.
A file checkpoint may undo an edit. It won’t unsend an email or recreate a live reservation by itself. Know which changes you could undo before giving the agent permission to try them.
What are you paying for when an agent makes a change?
You may be paying for a plan, a pool of AI usage or a separate account that pays as the AI works. That covers the work of changing the project. It doesn’t pay the bill for the app customers use.
Try a typical small change and check where it appeared in your account. If an agent is reading many files to answer a simple question, try pointing it to the part you’re working on. Ask what happens when the allowance runs out, especially whether work pauses or spending continues.
| Bill | What it pays for |
|---|---|
| Agent plan | Access to the editor or agent and any included allowance. |
| Extra AI use | Work beyond an allowance or work billed to an outside AI account. |
| Running app | Services such as hosting, customer accounts, saved records and emails. |
When is it worth moving work to the cloud or writing project rules?
Choose cloud work when you need the task to run on a separate computer. It needs its own copy of the project and access to whatever makes it run. Give it test settings deliberately; another computer can still connect to live services.
Write project instructions when you keep giving the same correction. A short note that says how to start the app, which database is for tests and how to check a booking can save time. Access settings still decide what the agent can reach. Neither a cloud session nor a rules file makes the first change easier to check by itself.