Help
How to get a first brief, what the settings do, and a fix for every message the app can show. General questions are in the FAQ.
Your first brief
- Add the material. Drop files on the card, click the paperclip to choose them, click the link tool to paste an address, or paste text straight into the box. Up to 20 sources and 10 links per run.
- Say what you want. Leave the box empty for a general summary, or write the task: what do I owe and by when, list the action items, compare these two, rewrite this plainly.
- Press the arrow. On a first run, the app may ask before downloading an on-device model and shows the size. Agree once; it is cached.
- Read the label. Under every brief: on your device (green) or cloud (blue). Copy the brief out if you want to keep it; nothing is saved for you.
Settings
The gear on the card opens two choices.
Engine
- Automatic (default): on your device when the browser can, otherwise the cloud. The result says which.
- Chrome built-in AI: only the browser's own model. Desktop Chrome and Edge 138 or newer.
- Private model on this device: only the downloaded model, over WebGPU. Works in Safari 26, Firefox, Chrome and Edge on most hardware.
- Cloud model: only the capped cloud path. The strongest read on a big pile; the text is relayed and dropped.
A pinned engine never falls back. If it cannot run, the app stops and tells you why, which is the setting to use when a file must not leave the device.
Length
Short is a paragraph. Medium suits most documents. Long is for contracts and reports where every clause matters. The task you type steers the content; length only sets how much of it you get.
Something is not working
Find the message or the symptom, then the fix.
"Chrome built-in AI is not available here"
Why: The built-in model needs desktop Chrome or Edge 138 or newer, a desktop operating system, about 22 GB of free disk space, and either a GPU with more than 4 GB of memory or 16 GB of RAM. Phones do not qualify yet.
Fix: Update the browser and free up disk space, then run again: Chrome downloads the model once and it works for every site afterwards. If the machine cannot meet the requirements, set the engine to Automatic and the private model or the cloud will take over.
"A private model needs WebGPU"
Why: The private on-device model runs on your graphics chip through WebGPU, which this browser has off or does not have.
Fix: Use current Chrome, Edge, Safari 26 or a recent Firefox. On an older browser, leave the engine on Automatic; the cloud fallback runs instead and the result says so.
The model download stalls or restarts
Why: The private model is about 1 GB on a computer and 350 MB on a phone. A dropped connection or a browser that clears site storage can lose the download.
Fix: Stay on the page until the progress reaches the end. The download is cached, so it happens once per browser. If storage is very low, the browser may evict it; free up space and run again.
"No readable text" on a PDF or image
Why: The file has no text layer and recognition found nothing: a blank scan, a photo too blurry or dark to read, or a PDF that is really a picture of a page at low resolution.
Fix: For a scan, try a straighter, sharper copy. For a photo, shoot square-on in good light and crop to the text. Password-protected PDFs also read as empty; remove the password first.
A link returns nothing or only a headline
Why: The page needs a login, sits behind a paywall, or draws its content with scripts after loading. The server can only read what a public page sends as text.
Fix: Open the page yourself, select the text, and paste it into the card. Pasted text is read entirely on your device.
"You have used today's cloud runs" or "at its daily limit"
Why: The cloud fallback is capped per person and per day so it can stay free.
Fix: Run on your device instead: desktop Chrome or Edge with the built-in model, or any WebGPU browser with the private model. There is no cap on your own hardware. Otherwise, try again tomorrow.
The run is slow
Why: First runs set up a model. Scanned pages and images each add a few seconds of recognition. Very long material is read in parts and then combined.
Fix: Let a first run finish; later ones are much faster. Crop images to the text. For a huge pile on a slow machine, the cloud model in settings is quicker.
The brief missed something or got a number wrong
Why: Small on-device models trade some precision for privacy and speed, and a poor scan can put the wrong digits in front of the model.
Fix: Ask for the specific thing by name ("every amount and its due date"), choose Long, or switch the engine to the cloud model for the strongest read. Always check amounts and dates against the original.
Settings do not stick between visits
Why: Settings are kept in the browser only while functional storage is allowed in the cookie notice, and a private or incognito window drops them on close.
Fix: Open Your privacy choices in the footer and allow Functional. Use a normal window if you want settings to persist.
Supported input
PDF (text or scanned), Word as .docx, plain text, Markdown, CSV and TSV, HTML, JSON, logs, subtitles, PNG, JPEG, WebP, GIF and BMP images, pasted text, and links to public pages. Not read: .doc, encrypted files, password-protected PDFs, and pages that need a login. Each kind has its own page: PDFs, Word documents, web pages, screenshots and images.
Still stuck?
Questions, bugs and ideas go to the contact address on the privacy page. Say which browser and which engine label you saw; the brief itself is never needed and never asked for.