Contents
The short version
- Claude first. Claude, by Anthropic, does most of the work. Auto picks the right size of model for each job, and you can always choose yourself.
- Never trained on your data. We only use providers that don't train on what we send, and we don't either.
- It asks before it acts. The AI proposes; you approve anything that sends, pays, books or deletes.
- It can be wrong. Check anything that matters.
- It helps you learn. It follows your course's AI rules and stays out of live quizzes and exams.
The models
- Claude by default. Reposance uses Anthropic's Claude models: a small one for quick jobs, a larger one for most work, and the largest for the hardest problems. Only Claude drives the cloud computer.
- Auto, or your choice. Auto chooses for each kind of work. In Settings › Models you can pin a model for a kind of work, or pick one in a chat. We never swap your choice for another company's model without telling you.
- Other models. Some lighter jobs may later use open models, such as OpenAI's gpt-oss, served by US hosts like Groq, but only after they pass our quality and safety checks. Other providers (such as OpenAI or Google) are used for a task only when that feature is on or you choose them.
- No Chinese-origin models. We don't use models from Chinese developers, through any host.
- Bring your own key. You can add your own API key for a provider in Settings › Models. That use goes on your bill with them instead of your credits. We keep the key encrypted with your own key and never show it again.
- No AI for small jobs. In Settings › Models you can let simple jobs run without a model at all.
What the model sees
For each step, the model gets what that step needs, and no more:
- your message, and the chat so far;
- the parts of your memory that matter to it;
- the mail, events, files or web pages the task is about.
It never gets your passwords, sign-in codes or bank transactions. Live voice goes straight from your browser to the speech provider, and Reposance doesn't keep the audio.
Training
- We don't train or evaluate AI models on your data.
- Neither do our providers. We only use providers whose terms forbid training on what we send. Any that train on it are refused.
- They may keep it briefly. Providers may keep what we send for up to 30 days to check for abuse, then delete it. Sensitive information goes only to providers that keep it 30 days or less.
Memory
Reposance remembers what helps you: your courses, your people, how you like things done. This makes it better for you without changing the model itself.
- You can see it. Settings › What I know shows everything it knows, and why it thinks so.
- You can fix it. Correct a fact, confirm or reject a guess, or forget an item, and what was learned from it goes too. When it notes something new, it says so, with Undo.
- You can stop it. Pause learning, download your memory, or forget everything (with your passkey).
- Private chats keep nothing at all.
Acting, with your approval
Reposance can write an email, add an event, fill in a form or use a browser for you. The AI model only proposes these steps. A separate gate, which isn't AI, checks each one, and anything that sends, pays, books, submits or deletes waits for your yes. Text from web pages, mail and files is treated as data, never as orders. More on Security & data use.
Mistakes and limits
- It can be wrong: wrong facts, wrong maths, outdated details, or a misread file. It shows its sources when it uses them, so you can check.
- It can miss context it was never given, such as a change your instructor announced in class.
- It isn't a professional. For medical, legal, financial, immigration or safety decisions it gives information, not advice. Talk to a qualified person.
- Sites can block it. When a site won't let it through, it shows you what it found and hands the step to you.
- Tell us when it's wrong. Correct it in the chat, or write to support@reposance.com.
Academic integrity and help modes
Reposance is built to help you learn, and to keep you on the right side of your course's rules. Each course has a help mode:
- Full help: worked solutions, with a note of how AI helped that you can hand in.
- Learn: hints, explanations and similar worked examples, but never the final answer to the exact graded question. This is the default when Reposance hasn't found the course's rule.
- No AI: nothing on graded work. Practice is still fine.
Reposance reads the mode from the course's syllabus, and when the wording is unclear the stricter reading wins. You can always choose a stricter mode yourself. While a live quiz or exam is open, that course gets no help at all, and Reposance never opens quiz or exam pages. Your school's rules come first: you're responsible for following them.
Teens
Reposance is for people 13 and over (14 and over in Quebec). From 13 to 17 the personal assistant works fully, while other people can reach you only in spaces you've been invited to. Being findable, messages from outside those spaces, paying and banks wait until 18. Reposance never makes sexual images involving minors, intimate images of real people, or realistic images of private individuals, for anyone.
Questions
Ask us anything about how Reposance uses AI: support@reposance.com. A person reads every message.