What can you ask a PDF?
- Summarize: a one-paragraph overview, then the main points as a list with page numbers. Works as an AI PDF summarizer for reports, papers and contracts.
- Ask questions in your own words (“What is the notice period?”, “Which figures changed since last year?”). Follow-ups keep the context of the conversation.
- Extract tables as Markdown tables you can paste into a spreadsheet or document.
- Pull out key fields: names and parties, dates and deadlines, amounts, reference numbers.
- Translate the whole document, keeping headings, lists and tables.
- Explain simply: plain-language explanations of legal, medical or technical text, with anything you need to act on.
Answers are written to stay with the document: the model is told to say when the PDF doesn’t contain the answer rather than guess, and to point to page numbers. Check anything important against the page it cites; models can still be wrong.
How it works with the editor
- In the editor, click Chat with PDF in the top bar. The panel opens on the right; the document resizes to fit.
- Open the settings (gear icon), choose a provider and paste your API key. The key field links to the provider’s key page.
- Save your document in PdfCraft with File › Save (Ctrl/⌘+S). It downloads as usual, and a copy appears in the panel. You can also upload a PDF or drop one on the panel.
- Pick a task (Summarize, Key points, Extract tables, Key fields, Translate, Explain simply) or type a question, then press Ask (or Ctrl/⌘+Enter).
- Copy an answer, download it as
.mdor.txt, or download the whole conversation.
Because the panel reads the saved file, it sees what you saved: rotate, remove or redact pages in PdfCraft, save, and the next question is about the new version.
Supported providers
| Provider | Default model | Get a key |
|---|---|---|
| OpenAI | gpt-6.1-sol | platform.openai.com |
| Google Gemini | gemini-3.8-flash | aistudio.google.com |
| OpenRouter | google/gemini-3.8-flash | openrouter.ai |
| Custom (OpenAI-compatible) | any model your API serves | your own endpoint |
OpenAI receives the PDF through its Responses API as a file input, Gemini as inline document data, and OpenRouter as a file part that it routes to models from OpenAI, Google, Anthropic, xAI and others. The custom option works with any service that implements OpenAI-style /chat/completions with file parts and accepts browser requests.
Privacy: what goes where
| Data | Where it goes |
|---|---|
| Your API key | Stays in your browser. Sent only to the provider, in the request header. |
| The PDF you pick and your questions | Sent from your browser straight to the provider you chose. OpenAI requests ask it not to store the response. |
| Everything else you do in PdfCraft | Never leaves your device. |
| pdfcraft.run servers | Serve the static editor files. They receive none of the above. |
The provider’s own terms and data policy apply to what you send them. For confidential documents, remove what the model doesn’t need first: in PdfCraft, extract only the relevant pages or redact names and numbers, save, and ask about that copy.
FAQ
Is Chat with PDF free?
pdfcraft.run charges nothing and adds no markup, because the request never passes through us. You pay your AI provider for what you use at their normal API prices. A short report costs a fraction of a cent with a fast model; a long book costs more, and every follow-up question sends the PDF again.
Where is my API key stored?
In your browser only. By default it lasts for the current tab (sessionStorage). Tick “Remember my key on this device” to keep it in localStorage instead, and use “Forget all keys” in the panel settings to delete it.
How large a PDF can I send?
Up to 30 MB. OpenAI accepts files under 50 MB per request and Gemini accepts inline PDFs up to 50 MB; the file is sent base64-encoded, which adds about a third, so 30 MB keeps the request inside those limits. Gemini reads up to 1,000 pages. For anything bigger, split the PDF in PdfCraft first (Pages › Split document).
Can it read scanned PDFs?
Usually, yes. Current OpenAI and Gemini models read each page as an image as well as its text, so they can read scans. Through OpenRouter, models that take files directly get the PDF as is; for text-only models the panel asks OpenRouter’s free Cloudflare parser to convert it to text first, which can miss scanned pages.
Which model should I pick?
For OpenAI, gpt-6.1-sol is a good balance for long documents and gpt-6-luna is the low-cost choice. For Google, gemini-3.8-flash is fast and handles long PDFs well. On OpenRouter, the model list in the panel shows only models that accept files. “Load models” lists everything your key can use.
Can the AI edit my PDF?
No. The panel reads the document and writes answers: summaries, tables, translations, explanations. Copy them or download them as Markdown or plain text. To change the PDF itself, use PdfCraft’s own tools in the editor.
Why does a request fail with a CORS or network error?
The custom provider option needs an API that accepts requests from browsers. OpenAI, Google Gemini and OpenRouter all do. Some self-hosted or regional gateways don’t; use one of the built-in providers instead.