The assistant and models

The assistant on the right runs on your computer and sees this project's data and code:

• Answer questions about your data: query tables, aggregate, draw charts;
• Build features: add pages, fields, buttons and schedules;
• Use what's on this computer: read your files, run command-line tools, use connected MCP servers;
• Fix page errors: when the back office shows an error, click "Ask the assistant to fix" at the top. Only then is the error sent to it;
• Run tasks: buttons that hand work to the assistant open a background chat; watch it on the right.

When it needs more information, it shows a few questions for you to pick or fill in, then carries on.

Settings → Models offers three kinds, and you can switch at the bottom of the chat any time:

• Your own provider: pick one, add an API key, tick the models you want. Keys stay on this computer;
• Local agent: Claude Code / Codex installed on this computer;
• creght platform models: available once creght is connected, billed in AI credits.

Token usage per model call is in Settings → Usage, recorded only on this computer.

• Instructions for the assistant: in Settings → Projects, write what you expect from the assistant in this project (style, things it must not do). It goes into the system prompt and wins over the project's own notes; template upgrades don't change it. Takes effect from the next new chat;
• Skills: manage instructions for the assistant in Settings → Skills. Only enabled skills enter its context.

• It can't read the values you store in Settings → Keys;
• It won't approve content for publishing on your behalf, or delete things on a platform, unless you explicitly ask.