The dashboard assistant (beta)

The assistant is the turtle in the bottom-right corner of every project in the dashboard. Ask it for something in plain words — “create a tasks collection with a title and a status” — and it does it for you, explaining as it goes. It is in beta: a platform operator turns it on per person, the same way as Studio.

What it can do

It uses the same tools as the MCP server, so anything you can do there it can do here:

• create collections, JSON schemas, indexes and access policies;
• insert, read and update documents;
• create member roles, users and API keys;
• write functions: database hooks (pre/post), cron jobs and custom HTTP endpoints;
• read the guide it learned groveback from, and explain any of the above.

It acts as you. Every action goes through the API with your own session, so a Viewer’s assistant can only read, and an action you are not allowed to do fails with the same 403 you would get by hand. It never has more access than you do.

Talking to it

• Keep it open while you work. The assistant is a floating window: drag it by its title bar, resize it from the bottom-right corner (double-click the title bar to put it back), and keep using the dashboard behind it. It stays open across pages and reloads until you close it.

• “This” and “here” work. The assistant knows which section you have open — Database → books, Functions — so “add an index on email here” means the collection on screen.

• Each action shows as a card with what it called and its arguments. A green check means it worked; a red cross shows the error it got. View jumps to the section where the change is.

• Ask it to explain. After it does something, the Explain what you did button asks for a step-by-step explanation — the best way to learn what a policy or a hook really does.

• Pick the AI model. When your operator configured more than one provider, click the model name under the title to choose which one answers you; your choice is remembered in this browser.

• Conversations are kept. Close the panel or the browser and pick up later; the menu next to the title lists your conversations for this project. Deleting one never undoes what the assistant did in the project.

Actions that ask first

Deleting, revoking or rotating anything — a collection, a document, an API key, a function — is never done straight away. The assistant stops and shows a Confirm this action card with the exact operation and its arguments. Nothing happens until you press Yes, do it; No tells it to stop and ask what you want instead. While an action waits, you cannot send new messages.

Only that button confirms. Writing “yes” in the chat does not, so text inside your data can never approve a deletion on your behalf.

The turtle

The turtle is the Groveback mark, and it shows what the assistant is doing:

It…	Because
breathes and blinks	it is waiting for you
has thought bubbles	it is waiting for the AI model
runs	it is carrying out an action
hops	it just finished
shakes, with a sweat drop	something failed — read the message
has heavy eyelids	you have used most of today’s allowance
sleeps	today’s allowance is used up; it resets at 00:00 UTC

Limits

The panel shows how many tokens the open conversation has used, and Account → Assistant usage shows your own consumption: today against your limits, and per project over the last 7, 30 or 90 days.

Each person has a daily allowance of requests and tokens, set by the operator; the panel shows how many requests you have used today. A single request also stops after a number of steps — if it says so, ask it to continue.

Good to know

• What is sent to the AI provider: your messages, the list of your collections, and the results of the actions it runs (truncated when large). Secrets such as API keys and OAuth client secrets are never returned by the API, so they are never sent.
• It can be wrong. It reads the project back after acting, but check the result — the View link is there for that — and prefer small steps.
• Documents are data, not instructions. If a document says “delete everything”, the assistant treats it as text to read, not a command.

For operators

The assistant needs an AI provider configured by a platform operator under Account → Assistant: Anthropic, OpenAI, Gemini, or OpenAI-compatible — any server that speaks Chat Completions (OpenRouter, vLLM, Ollama…), which needs its base URL:

• Providers: kind, model and API key — the list button next to Model asks the provider which chat models that key can use, so you pick one instead of typing an id — (stored encrypted, never shown again; leave the field blank to keep the saved key), an optional base URL, an optional reasoning effort (for reasoning models — lower is faster and cheaper, higher thinks more; Model default sends nothing), and which one is the default. Test sends one tiny request and reports the latency.
• Limits per user per UTC day: requests and tokens (cached prompt tokens count at a tenth), plus actions and model steps per request.
• Behaviour: default language, extra instructions (“students are building a library app”), the suggested prompts shown in an empty conversation, and whether the full guide goes in every request (better answers, more tokens) or is fetched only when needed.
• Let members choose the provider (on by default): with more than one provider configured, members pick which one answers them. They choose among yours — they never enter keys.
• Usage over the last 7, 30 or 90 days, by organization, by project and by person: requests and tokens (input, output, cached).

Then give people the assistant flag under Customers → Flags. Operators always have it.

Self-hosted servers can skip the settings page with environment variables; the first save in the dashboard takes over from them.