Watch your AI agent use a phone, live.
Live view streams the phone’s screen while your agent works and draws every action on it: the cursor moves to each target, taps ripple, swipes leave a trail and typed text appears. Each step gets a caption with its time, so you confirm the whole run by eye.
One link, from your agent, your terminal or the dashboard.
A real run, at real speed.
An agent looks up the Matterhorn on Wikipedia over MCP. The cursor, the trails, the typed text and the log are the live view’s own.
Every step, confirmed by eye.
Nothing in the view depends on sound. People who are deaf or hard of hearing, or who simply work with the sound off, see each action where it happens and read what it did.
On the screen
- A cursor that glides to each target
- A ripple for every tap, a trail for every swipe
- The text as it’s typed, dots for passwords
In the log
- One caption per step, with its time
- A Now line for the step in progress
- Failed steps marked, with the reason
For every viewer
- Captions announced by screen readers
- Reduced motion respected
- Fits a narrow window or a second screen
Open it in one step.
Over MCP, ask your agent to show you the phone: it calls watch_phone and hands you the link. From a terminal, one command opens the view in its own window.
# Your terminal: the phone opens in its own window
npx mobilebox watch <phone_id>
# Your code: a link to show or open
curl -X POST "https://mobilebox.dev/v1/phones/<phone_id>/view" \
-H "Authorization: Bearer $MOBILESANDBOX_API_KEY"Questions
Opening, sharing and following a live view.
What is the live view?
A page that streams a Mobilebox phone’s screen while your agent uses it and draws each action on top: the cursor, taps, swipes and typed text, with a caption and time for every step.
How do I open it?
Ask your agent: over MCP it calls watch_phone, over REST POST /v1/phones/{id}/view, and either returns the link. From a terminal, npx mobilebox watch <phone_id> opens it in its own window. In the dashboard, every running phone has a Watch button.
Can people who are deaf or hard of hearing follow along?
Yes. Every action is drawn where it happens and written as a caption with its time, so nothing relies on sound. Screen readers announce each caption, and reduced-motion settings are respected.
Can someone watching control the phone?
No. The view is read-only: it shows one phone’s screen and its captions, and sends nothing back. Each link is short-lived; ask for a fresh one any time.
Are passwords shown?
No. Text typed into a password field shows as dots, masked before it leaves our servers.
Does watching cost anything?
No. The view adds nothing to the phone’s per-minute price, and watching never keeps a phone running. See pricing.
Read more
A successful tap is not a successful task
Design Android agent checks around outcomes, screen evidence, failure traces, and safe cleanup.
Give your coding agent an Android sandbox over MCP
Add the remote MCP server to Claude Code or Cursor, see every phone tool, run a bounded first task and know what is proven.