AI Studio

Studio is the full coding agent, opened on one deployment. Where the [AI Chat](ai-chat.md) drawer answers questions next to whatever page you are on, Studio gives the agent a workspace: it reads your project, writes the changes, installs what is missing, restarts the app and checks that it actually answers.

Opening Studio

Open a deployment, then use the Studio button. Studio always works on that one deployment.

The workspace

Studio is a single screen split into three panels you can show, hide and resize:

• Agent: the conversation. You describe what you want, the agent works and reports what it did.
• Files and environment: the file tree and the environment variables of the deployment, so you can see what the agent sees.
• Console: the live console, with the power controls (start, stop, restart) and the resource usage of the deployment.

The top bar shows the live state of the deployment: running, starting, installing, stopping or offline.

How a task runs

A typical turn goes: understand the request, read the relevant files, plan, edit, install any missing package, restart the deployment, then verify. Verification is real: the agent reads the console and, for a web app, requests the public URL to see what it answers.

You stay in control. The agent asks before anything irreversible and reports each file it touched.

Checkpoints and undo

Every turn records a checkpoint of the files it changed.

• Click a file in the turn to open a diff, so you see exactly what changed.
• Restore brings a file, or the whole turn, back to how it was before.

This is the safety net: you can let the agent try something, look at the diff, and roll it back in one click if you do not like it.

History

Each deployment keeps its own Studio threads. Reopening Studio lands on the most recent one, and older threads stay available, so you can pick up a task days later with its full context.

What you need to use it

Studio needs one of:

• purchased AI Tokens in your wallet (any pack), or
• your own provider key, activated on the AI assistant page.

The free weekly allowance does not open Studio. It is there to try the chat assistant, not to run a coding agent.

What it costs

A finished task typically costs between 20,000 and 30,000 AI Tokens, so the smallest pack covers roughly a hundred of them. You are billed for what the agent reads and writes. Context it has already seen counts as a tenth, which is why a long session stays affordable.

With your own provider key, nothing is spent from your AI Token balance.

Getting good results

• Point at the file or the feature. "Fix the slash command registration in commands/ping.js" beats "my bot is broken".
• Let it read the error. If the app crashed, say so and let the agent open the console itself.
• Review the diff before you move on. It is faster than debugging a change you did not notice.
• Keep one thread per subject. A fresh thread is cheaper and sharper than a long one that drifted.