Getting started

Signing in creates your account and adds $0.50of credit — enough for a real conversation on a good model. Pick a model from the composer, type, send. The reply shows its exact cost underneath, and a meter above the composer counts up while it arrives.

Switching models mid-conversation keeps the whole transcript. Use a cheap fast model to explore and a frontier one for the part that matters; nothing is lost between them.

A persona is a saved set of instructions. Studio ships six — Straight answers, Devil’s advocate, Analyst, Writing partner, Code reviewer, Socratic tutor — and you can write your own from Personas in the sidebar.

Apply one from the composer, next to the model picker. Applying it copies its text into that chat, so you can then edit it freely; editing marks the chat “Custom” and leaves the persona itself untouched. This is deliberate — changing a persona should never rewrite the instructions a finished conversation was actually held under.

A persona can also carry a model. “Code reviewer” on a frontier model and on a cheap one are different tools, and the persona is allowed to have an opinion about which.

The single most useful thing you can write into one is how to disagree with you. Most prompts say how to answer and leave that out, which is why most assistants fold.

Every chat can carry a goal: one sentence about what you are trying to achieve. The model keeps it in view and maintains a running plan alongside the conversation, which you can open from the header.

It matters most on long chats, where the thing you actually wanted scrolled off the top an hour ago.

Upload documents from Knowledgeand they apply to every chat. Upload them from inside a chat’s instructions panel and they apply only there, keeping a reserved slot in the reference material even when your global library is large.

Text is extracted, split and indexed. The original file is not stored. Supported: PDF, Word, Markdown and plain text.

What to put in: the things you have already worked out. Your notes on a subject, your style guide, your architecture decisions, the papers you are working from. Not a dump of everything you own — retrieval gets worse, not better, when the index is full of things you never ask about.

Add a key from API keys in the sidebar, then switch a chat to Use my own key in its instructions panel. It is per chat, not per account, so you can keep some work on your key and some on Studio credit.

Which key pays, in order:

1. A vendor key matching the model — an openai/* model goes straight to OpenAI, skipping the gateway margin entirely.
2. Your OpenRouter key, which covers the whole catalogue.
3. Studio credit, if you have connected nothing.

Falling through to the third is deliberate: removing a key should make chats cheaper, not broken. Keys are verified against the provider’s own endpoint before they are stored, so a typo fails on the settings page rather than mid-chat.

A coding agent with a real container per project. Clone a repository or upload a zip, then work in Ask, Research, Plan or Build mode — which are four different sets of tools, not four different instructions.

Ask is the one for the middle of a task: it reads your project and answers questions about it — where a setting lives, what a value should be — and it is the only mode that cannot change your plan as well as your code, so stopping to ask never costs you your place in the work.

Start in Plan with approvals on Manual. Once you trust what it is doing on a given project, move to Auto; it still stops and asks before anything destructive.

Your project’s files live in a container on our machine, and a commit the assistant makes stays there until you send it somewhere. Publishing to GitHub and downloading a zip are both one button, and Git in Studio Code is the page that explains which one you want.

The full explanation

Every message shows what it cost. Usage rolls that up by day, model and project, and Credits in the sidebar is your balance. New accounts have a daily ceiling; you can raise or remove it once you have topped up.

The two things that actually drive a bill are the model you picked and how much history is being re-sent each turn. Branching is the cheap fix for the second: fork a long chat rather than carrying forty messages of irrelevant context into every new question.

Write to garv@gdautomate.com. Include roughly when it happened — everything is timestamped, which makes most questions quick to answer.