Docs · v1.0.0

Set up once, then forget it.

Two minutes to a working setup, whether you are using a cloud key or a model on your own machine.

Install

Unprompter needs Chrome 116 or newer, or any Chromium browser of the same vintage — Edge, Brave, Opera, Arc.

Add it to your browser

Install from the Chrome Web Store, or load the unpacked build from chrome://extensions with Developer mode on.

The settings page opens itself

On first install. Nothing works until a model is connected, so this is the one step you cannot skip.

Connect a cloud key

Pick your provider, paste the key, press Test this setup. The test sends a 2×2 pixel image and checks the whole chain — key, endpoint, model, and whether the reply parses — for a fraction of a penny.

OpenAI — a key from platform.openai.com. gpt-5-mini is the value option.
Anthropic — a key from console.anthropic.com. claude-sonnet-5 reads images most carefully.
Google — a key from aistudio.google.com. gemini-2.5-flash is fast and cheap.
OpenRouter — one key, any model behind it.

Chrome asks for permission to talk to your provider's domain the first time. Decline it and the call cannot be made — that prompt is the browser's, not ours, and there is no way around it.

Connect a local model

Free, private, and unlimited. You need a vision model — a text-only model will return nothing useful.

Ollama, from a terminal

# a vision model, about 6GB
ollama pull qwen3-vl:8b

# let browser extensions talk to it, then restart Ollama
launchctl setenv OLLAMA_ORIGINS "chrome-extension://*"   # macOS
export OLLAMA_ORIGINS="chrome-extension://*"             # Linux / Windows WSL

Then choose Ollama in the settings, set the model to qwen3-vl:8b, and test. LM Studio works the same way: switch on its local server and point Unprompter at http://localhost:1234/v1/chat/completions.

The first reading on a local model is slow — a minute or more while the model loads into memory. Every one after that is quick. Unprompter waits three minutes before giving up on a local model, and ninety seconds on a cloud one.

Activate Pro

Paste the key from your order email into the settings page and press Activate. It claims one of your 3 browser seats and turns on every format, the variations, batch, compare and export.

Moving machine? Release the old seat from the manage page or from the settings page in that browser. A released browser drops back to the free tier rather than breaking.

Reading an image

Right-click any image → Unprompt this image.
Hover an image over 128px and click the badge in its corner.
Drag a region for canvases, video frames, maps and CSS backgrounds.
Drop, paste or pick a file straight into the panel.
Every image on the page (Pro) for moodboards and galleries.

The colour palette appears before the model has answered — it is measured in your browser. Click a swatch to copy its hex.

Shortcuts

Alt+Shift+U — drag a screen region.
Alt+Shift+P — open the panel.
Esc — cancel a region selection.
Change either shortcut at chrome://extensions/shortcuts.

Modes and the faithfulness dial

Everything reads subject and style together. Style only deliberately ignores who or what is in the picture — the right choice when you want a look for your own subject, and the polite choice when the picture contains a person. Subject only does the reverse.

The dial (Pro) decides how much of the reading reaches the prompt. Verbatim uses every field for a close recreation; loose keeps the subject and the mood and drops the forensic camera detail, which gives a generator room to be interesting.

Your library

Everything you read is kept in this browser — searchable across every field and every generated prompt, so "teal" finds the shot whether teal came from the palette or the grade. Star the ones worth keeping: favourites survive when older entries are trimmed. Pro exports the lot as JSON, CSV or Markdown.

Troubleshooting

"Could not reach Ollama"

Nine times in ten this is OLLAMA_ORIGINS. Set it to include chrome-extension://* and restart Ollama — setting it in a shell that Ollama was not started from does nothing.

"Could not read that image"

The page will not hand the file over and you declined the site permission. Either allow it when asked, or drag a region round the image instead — that route needs no permission at all.

The model returned no JSON

Usually a text-only model being asked to look at a picture. Check the model name is a vision model. On Ollama, llava, llama3.2-vision and qwen3-vl all work.

"All three browser seats are in use"

Release one from the manage page using your key. Reinstalling the extension in the same browser counts as a new browser, because the device id is regenerated.

Still stuck: hello@ulogon.com.