Draft
Reply in your words
Tell it what to say — “politely decline, propose next Tuesday” — and pick a tone. It answers in the language of the email, so a Spanish thread gets a Spanish draft. You always read the draft before it goes anywhere.
No server · No account · No telemetry
emailito lives inside Gmail and Outlook. It drafts replies and summarizes long threads using a model that runs on your own GPU — so the email you are reading is never uploaded, logged, or remembered.
Free and open source. Works in Chrome and Edge.
Where does your email actually go?
↓ leaves your machine ↓
Nothing crosses the line. The model runs inside the browser on your GPU. No request carries your email anywhere.
What it does
Draft
Tell it what to say — “politely decline, propose next Tuesday” — and pick a tone. It answers in the language of the email, so a Spanish thread gets a Spanish draft. You always read the draft before it goes anywhere.
Summarize
An overview, the key points, and who owes what — in the thread’s own language. The private alternative to the AI summaries built into your inbox.
Insert
One click puts the draft in the compose box — opening the reply for you if none is open — or copies it if you would rather paste it yourself.
Engines
Two engines keep everything on your computer. The third is better at writing, and we tell you exactly what it costs you.
A compact model runs inside Chrome on your GPU. First use downloads it once and caches it; after that, drafts start instantly and work with no network at all.
Needs a WebGPU-capable machine with ~3 GB free VRAM · ~2 GB download, once · smaller and larger models selectable in settings
Stays on your machine
Already run Ollama? Point emailito at it for bigger, better local models. The extension refuses any Ollama URL that is not loopback, so this cannot quietly become a remote call.
Needs the free Ollama app on the same computer · any model you have pulled
Stays on your machine
The best drafts, using your own Anthropic API key. Your email goes straight from your browser to Anthropic — never through a server of ours, because there isn’t one. You see a full disclosure before the first request.
Needs an Anthropic API key with credit at console.anthropic.com · a separate balance from a Claude.ai subscription
Sent to Anthropic
Not sure which model to run? Choosing a model has the VRAM figures for all six on-device models, why summaries punish small models more than replies do, and how to find out what your GPU can take.
Privacy
“It stays on your machine” is easy to write on a website. In emailito it is a line in the extension’s manifest that Chrome itself enforces — the complete list of hosts the extension is permitted to contact:
// manifest.json — content_security_policy connect-src 'self' https://api.anthropic.com // Claude mode only https://huggingface.co // model weights https://*.hf.co // (data, not code) http://localhost:* // your Ollama http://127.0.0.1:*
In the inbox