AI Use

docraft uses an AI agent to write and typeset your documents and to carry out your annotations. This page explains how that works and what it means for you.

Which models we use

The agent runs on Claude models from Anthropic, accessed through Anthropic’s commercial API. By default it uses a fast, inexpensive model; on paid plans, hard tasks can switch automatically to a stronger model while enough points remain. We may change the models we use as better ones become available.

What is sent to the model

For each turn, the agent receives what it needs to do the work: your request and chat messages, the current document, your annotations and checklist answers, and the attachments of that document (images may be re-encoded first). It does not receive other documents, other users’ data or anything from your account beyond the document you’re working on. The agent has no internet access; reference lookups by DOI are made by our server when you ask for them.

Your data is not used for training

• We do not use your content to train AI models.
• Anthropic does not use inputs and outputs from its commercial API to train its models by default, and deletes them from its systems within 30 days, except where longer retention is needed to enforce its usage policy or required by law.

See the Privacy Policy for how we store and protect your data.

Points and metering

• Points measure the model work each turn uses. A turn is charged when it ends, by what it actually used, including turns you stop or that fail.
• Before each turn we check your balance. If points run out mid-turn, the agent stops before its next step and keeps what it already changed as a version.
• Editing text, equations and chart data directly doesn’t involve the agent and uses no points.
• Plans, points and top-ups are listed on the Pricing page; your balance and usage per document are on the billing page in the app.

How we keep the output in check

• The agent asks you about anything that would otherwise make the content wrong, rather than guessing.
• Layout problems such as page count, overflow and overlaps are checked by code before delivery, not by the model’s own judgement.
• Every delivery comes with a checklist of what you need to confirm, every change is saved as a version, and changes outside the place you pointed at are flagged so you can undo the whole turn.
• Music and game records are engraved and drawn by dedicated programs from your files; the agent doesn’t compose notes or analyse games.

What you are responsible for

AI can make mistakes: in formulas, numbers, citations or wording. Check your document before you rely on it, especially for exams, publications or applications. You are responsible for the content you publish, and for having the right to use the material you upload. The agent cites only references in your document’s reference library; if you ask for one that isn’t there, it asks you to add it instead of inventing it.

Questions

Write to support@docraft.art.