Overview & result
jev-mobile is a durable Android sub-agent. It accepts a high-level task, stores it in SQLite, and runs an observe, normalize, decide, mutate, verify loop against a USB-connected device. Jev receives concise semantics and the technically valid actions and picks among them; it never generates coordinates, MCP calls or code. The README limits the validated scope to notes, checklists and text input.
From the original post
Fast structured Android control loops with TypeSafe Jev and Mobile MCP
How Jev fits in the loop
- A high-level task is stored in SQLite and picked up by one durable worker.
- The worker observes the Android UI tree and normalises it into a short list of valid actions.
- Jev picks one of those prevalidated candidates; ESCALATE is always available.
- The worker maps the choice to the device through an adapter, then verifies the result before the next step.
How to reproduce
- Inspect the original showcase and source material at https://github.com/Friedjof/jev-mobile
- Verify the bounded prompt and input candidate schema configured for Jev
- Benchmark decision latency and classification accuracy against baseline models
Why this build matters
Jev never sees raw device commands or coordinates, so invalid actions and stale references are rejected by construction instead of being caught afterwards.
Limitations
- Hardware validation covers notes, checklists, text input, crash recovery and persistence checks.
- Payments, account changes and other sensitive operations are out of scope.
- Low confidence, a small margin between the top two options, loops, timeouts and step limits stop the automation.