Fogar answers from a language model that runs inside your browser. No account, no key, no server. When you want a bigger brain, bring your own key and it uses that instead. Either way the choice is yours, and visible.
Free. Open source. One model download, 533 MB to 2.7 GB, sized to your machine and cached for good.
Most "private AI" claims are a paragraph in a policy. This one is a table you can verify in the Network panel.
| Mode | What leaves your browser | Where it goes | Cost |
|---|---|---|---|
| On this device | 0 requests once the model is cachedThe model file is fetched once from Hugging Face. Your questions never go anywhere. | Nowhere | Free |
| Cloud, your key | Your prompt and your keySent only to the OpenAI-compatible endpoint you typed. OpenAI, OpenRouter, Groq, or Ollama on localhost. | The address you chose | Your provider's rates |
| Web grounding | Your question, as a search queryOff until you tick “Search the web first”. Then DuckDuckGo’s instant answers and Wikipedia, with no key, or full web results with your own Brave Search or Tavily key. The model answers from what comes back, with citations. | DuckDuckGo and Wikipedia, or the search API you chose | Free |
No analytics. No telemetry. Nothing to opt out of.
Draft and rewrite, reply to a message, summarize, explain simply. Make your own with a small form builder, then share it as a link. Recipes are data, never code.
Plain words in, a confirmed time out. Fires as a Chrome notification. One click adds it to Google Calendar, with no account connection needed.
Paste your calendar’s private iCal address. Google, Outlook, iCloud, anything. Recurring events included, no account connection.
Todos, links, notes, weather with no API key, a favorite recipe. Order them, hide them, or hit Focus and keep only the ask box.
Start typing in the ask box and matching bookmarks appear underneath. Open one, or remove it.
Select text on any page, right-click, and get an explanation in a new tab. The page comes along: its title, address, and the text around your selection, so the model answers from what you were reading.
One menu on every answer: search the web and answer again with sources, let the model think longer, or ask your cloud model the same question. The better answer replaces the first one in the conversation.
A small local model is a good writer and a poor encyclopedia. Tick “Search the web first” for the questions where that matters: DuckDuckGo’s instant answers and Wikipedia with no key, or full web results with your own Brave Search or Tavily key. Sources are shown. Arithmetic is computed, not guessed.
A recipe declares its inputs and a template. Fogar draws the form, fills the template, and hands it to whichever model you are using. Because a recipe is plain JSON, you can paste one to a friend, post it on a blog, or open a share link and add it in one click. Copy the one below and paste it into Recipes → Import to see.
Rewriting is the thing a small local model does best, so the first built-in recipe is for exactly that.
{
"name": "Draft & rewrite",
"inputs": [
{ "key": "text", "label": "Your text", "type": "textarea" },
{ "key": "tone", "label": "Tone", "type": "select",
"options": ["Neutral", "Friendly", "Formal", "Direct"] },
{ "key": "length", "label": "Length", "type": "select",
"options": ["Shorter", "About the same", "Longer"] }
],
"template": "Rewrite the text below.\nTone: {{tone}}. Length: {{length}}.\n\n{{text}}"
}
On this device, cloud with your own key, or just a search box for now. You can change your mind any time.
Qwen3.5, quantized, from Hugging Face, in the size your machine can carry: 0.8B, 2B, or 4B. It lands in your browser's private file storage and never downloads again.
533 MB · 1.3 GB · 2.7 GBThe model wakes from disk, runs on your GPU through WebGPU or on your CPU across all its cores, and starts answering. Follow-ups keep the thread.
warm start 1.3 to 3.2 s · 44 to 92 tok/s · Apple silicon, WebGPUBuilt by Dylan Roy, who also makes UseSQL, TickerPal, and Kumkuat.
Yes, in local mode, once the model is cached. Turn off Wi-Fi and it keeps answering. Web grounding and cloud mode need a connection, by definition.
Depends on the size. In our tests the 0.8B and 2B models rewrite and summarize well but stumble on facts and arithmetic. The 4B gets the president, the percentage, and the sorting task right and still runs at 44 tokens a second on an Apple laptop. Fogar recommends the largest size your machine can run well, computes arithmetic itself, and offers web grounding for facts.
In Chrome's extension storage on this profile, and it is sent only to the endpoint you typed. Fogar has no server that could see it.
Storage, alarms and notifications for reminders, bookmarks for search, context menus for right-click, favicons for site icons, and access to huggingface.co to fetch the model. Cloud endpoints, search keys, and calendar addresses are requested one at a time, when you add them. The full list with reasons is in the privacy policy.
The model loads only on the new tab page and is released when you navigate away. Loading takes about a second from disk. On a machine without WebGPU it falls back to the CPU and is slower but works.
Chrome and Chromium browsers first. Firefox builds, but its extension pages lack the shared-memory feature the multi-threaded engine needs, so it would be much slower. Not shipping it half-baked.
Fogar is on its way to the Chrome Web Store. Until the listing is live, it installs from source in two commands and a drag-and-drop. This page will update the moment the store link exists.
git clone https://github.com/dylanroy/fogar && cd fogar npm install npm run build # chrome://extensions → Developer mode → Load unpacked # → pick .output/chrome-mv3, then open a new tab