MCP server
Works with any MCP client or your own agent. Start a task, watch the screen, answer questions, stop.
"openmob": {
"url": "https://openmob.fly.dev/mcp",
"headers": {"Authorization": "Bearer <your API key>"}
}Invite-only beta · Android 11 and newer
Give any AI agent a real Android phone.
One MCP server and REST API between your agent and a real phone. The phone judges every tap itself, so your agent can act without being handed the keys.
agent →
“Reorder my usual coffee beans and pay with the saved card.”
your phone →
- ALLOWED
- open Shop · search “house blend”
- ALLOWED
- tap “Add to cart”
- ASKED YOU
- use the saved card → approve with your fingerprint or PIN
- ASKED YOU
- tap “Place order · $18.40” → approve on the phone
- BLOCKED
- open Banking: banking apps are off-limits
Every tap
judged on the phone itself
Your model
pick the AI model for each task
For developers
Three ways in. One safety model.
Whichever way your agent connects, the same phone-side checks apply. Your key can tighten them, never loosen them past what the phone’s owner allows.
REST API and webhooks
Start tasks, stream events, answer questions. Scoped API keys and signed webhooks.
curl -X POST https://openmob.fly.dev/v1/devices/$DEVICE_ID/tasks \
-H "Authorization: Bearer $OPENMOB_API_KEY" \
-H "Content-Type: application/json" \
-d '{"instruction": "open YouTube and search for lo-fi"}'On the phone itself
No code: chat with your agent in the OpenMob app, pick a model, and watch it work. A live view and Stop are always one tap away on the dashboard.
Security
The server could be compromised. The phone is built as if it were.
Most agent tools trust whoever sends the command. OpenMob puts the final decision on your phone, with its own rules and its own keys, so a hacked server or a prompt-injected agent can at most ask: risky steps still wait for you, and blocked ones are refused.
- Your agent
- OpenMob server (not trusted by the phone)
- Your phone decides
- Your apps
Judged on the phone
Every action is checked on the device before it happens. Payments always ask you first; sends, posts, deletes and installs ask too, unless they’re what you asked for.
Keys that stay put
The keys that identify your phone and sign your approvals are sealed on the phone. They’re never backed up or moved to a new phone: a new phone pairs again.
Approvals happen on the phone
Purchases and other high-risk steps take your PIN, pattern or fingerprint. Approval notifications never show what’s being asked: you see it in the app.
Off-limits by default
Banking, payment, wallet and password apps are off-limits. OpenMob never types into a password field and never unlocks your phone.
Kept out of the record
Card numbers and Social Security numbers are hidden in the activity feed, history, API events and webhooks, and typing one always asks you. Screenshots are never stored.
Stop means stop
Stop from the dashboard, the API or the phone and the task stops right away. Disconnect on the phone and agents can’t reach it.
How it works
Paired in a minute. In control the whole time.
Pair your phone
Install the app, then scan a QR code or type a short code. You confirm on the phone with your screen lock.
Connect your agent
Make a scoped API key and add OpenMob to your MCP client, call the REST API, or just type in the app.
Watch, approve, stop
Follow every step in a live view. Approve risky ones on the phone. Stop any time.
Who sees your screen
To decide each step, what’s on screen is sent to the AI provider that runs your chosen model and, when the server uses one, as text to a screen-classification service. With Faster mode on (it’s off by default, and only you can turn it on, from the phone), small screenshots of screens without readable text, such as games, go to that service too. OpenMob never stores screenshots.
Invite-only, for now.
We’re opening OpenMob to a small group first. Have an invite link? Open it, or log in and choose “Have an invite code?”.