Local-first by default.
The model runs on your device. Your inbox, calendar, and chats never touch our servers. Cloud is opt-in, with your keys, your bills.
Local-first by design. Runs Gemma 4 on your phone. Plugs into Gmail, Slack, Notion, GitHub, and Calendar. Bring your own model when you need a bigger brain.

Most AI on your phone is a chat box pointed at a cloud. It doesn't read your email. It doesn't see your calendar. It can't post to Slack. It forgets you the moment you close the tab.
Xome is different by design. The model runs on your device. It connects to the apps you actually use. It takes action with your permission, in the foreground or while you sleep. And when a task needs a bigger brain, you bring your own — Claude, GPT, or Gemini, with your key, on your terms.
"It's the first AI on my phone that's worth keeping the icon on the home screen."
03 — HOW IT THINKS
Every turn is a small loop. The agent plans the next step, picks a tool, calls it, looks at the result, and decides whether to keep going.
No magic. No hidden waterfall of LLM calls. Just a tight loop you can pause, audit, and approve at every write.
When a step needs a bigger model, the loop seamlessly hops to your cloud key and back — your data never sits on our servers.
04 — FEATURES
The model runs on your device. Your inbox, calendar, and chats never touch our servers. Cloud is opt-in, with your keys, your bills.
Plug in Claude, GPT, or Gemini for harder tasks. Your key, your provider, your control.
Gmail, Calendar, Slack, Notion, GitHub. Real OAuth, real scopes, real writes.
Read freely. Write only with consent. You see the tool, the args, and a one-tap decline.
Point it at any MCP server and its tools become Xome's tools. Open ecosystem.
Talk to it like you'd talk to a colleague. On-device speech, no audio leaves the phone.
Triggers fire while you sleep. Push wakes the agent. Notification lands when work is done.
Snap a whiteboard. Get a structured summary. Local model, no upload.
Easy task → local Gemma. Hard task → your cloud model. The router picks per turn.
Self-host the relay. Read every tool definition. Ship your own MCP server.
Conversations stored encrypted with a key in your device's secure enclave.
One Flutter codebase. Native feel. No Electron, no React Native bridge.
05 — CONNECTS TO
// and anything that speaks MCP →06 — PRIVACY
The model that reads your inbox lives on your phone. Not on a server we run, not on a server anyone runs. The data path stops at your device.
When a task needs cloud, you bring your own key. Claude, GPT, Gemini — your provider, your bills, your audit log.
We don't keep a copy. We never see your messages. The relay that wakes background tasks carries nothing about you except a push token.
servers we run in the data path
local model by default
your bills · your control
07 — UNDER THE HOOD
All of it ships in one Flutter app. None of it is magic. All of it is open.
08 — WHERE THIS IS HEADED
Flutter scaffold, design system, model bench harness.
Plan/Call/Reflect loop, provider abstraction, secure storage.
First integration end-to-end with read and write scopes.
Four integrations rolling out, full read + scoped write.
Time, location, contacts, files. Real on-device.
Speech in/out, image input — all local where possible.
Point Xome at any MCP server, gain its tools instantly.
Triggers, schedules, push wake, audit log.
09 — COMMON QUESTIONS
The app and the local model are free. Cloud models cost what you pay your provider; we never see your bills.
No. The model on your phone reads them. We don't have a server with your inbox or your DMs.
Gemma 4 E4B by default. Switch to E2B for older devices, or FunctionGemma 270M for fast tool routing.
Use the cloud path with your own API key, or switch to the lighter E2B variant.
Yes, with admin-approved OAuth. We don't ask for any scope you wouldn't grant a normal client app.
Yes. Xome speaks MCP. Any MCP server you point it at becomes available as tools.
iOS background execution is constrained by Apple. We use silent push from a tiny relay we run, with periodic fallback. Real-world reliability is high but not 100%; we're honest about it.
Open beta opening soon. Join the waitlist for the invite — we're inviting in batches based on signup order.
we'll only email when it's ready · and once when it ships