Android AI agents that operate your real phone apps, with you in control
An Android AI agent that controls your phone: it reads the screen, taps, types and swipes in allowed apps without an API, and asks before any write.
A mobile AI agent observes the current Android screen, decides the next bounded action, taps, types, or swipes, then verifies the result. Melaya Device Control runs on a paired real phone, limits the agent to allowed apps, and pauses for human approval before publishing or other protected writes.
Computer use, but for the phone
Many important workflows end inside an app that has no practical API. Mobile agents use the visible interface, so they can work with the same controls a person sees. That reaches mobile-only and integration-poor apps, but it also means the workflow must handle screen changes, interruptions, permissions, and ambiguous UI state.
Melaya separates reasoning from device execution. A chosen model can propose steps, while Device Control performs bounded gestures on the paired Android phone and records what happened.
The safe mobile-agent loop
A reliable loop is observe, decide, act, verify, and stop. The agent reads the current screen, chooses one permitted action, performs it, and confirms the expected state before continuing. It should not assume a tap succeeded or that the next screen matches a memorized sequence.
- Per-app allowlist
- No root and no jailbreak requirement
- Approval before protected writes
- Replayable action and decision history
- Explicit timeout, retry, and return-to-Melaya behavior
Run your Android AI agent on a local model
A local AI agent for Android works today. Run Ollama or LM Studio on your own computer, connect the Melaya runner, and pick that model for the device agent. The model reasons on your machine, the phone performs the taps and swipes, and no cloud model provider sees your screen.
The model does not run inside the phone. Commands still travel through your Melaya account to the paired device, and small local models handle short, well-scoped flows better than long multi-app tasks. Local models are available on every plan, including the free Sandbox.
Where mobile agents fit
Good candidates include repetitive navigation, draft preparation, cross-app data entry, exploratory Android testing, and workflows blocked by missing APIs. Poor candidates include unsupervised financial actions, safety-critical device control, CAPTCHA bypass, or any workflow that violates an app's rules or a person's consent.
Melaya currently targets Android. It should not be evaluated as an iPhone-control product or as a replacement for a large device-farm testing service.
Frequently asked questions
What is a mobile AI agent?
It is an AI agent that can observe and interact with a mobile interface, usually by reading screen state and performing gestures such as taps, typing, and swipes.
Does Melaya need an API for every app?
No. Device Control can operate the visible Android interface, which is useful when a suitable API or connector is unavailable.
Does Melaya control iPhones?
No iPhone control is claimed. The current Device Control product is for Android.
Can the agent publish without approval?
Protected publishing actions are intercepted and staged for human approval. The user can review or edit the draft before it is submitted.
Can I run a local AI agent for Android?
Yes. Pick a local model served by Ollama or LM Studio through the Melaya runner on your computer. The model runs on your machine and the paired phone carries out the actions.
Which AI can control my phone?
With Melaya Device Control, the model you connect can: Claude, GPT, Gemini, a Claude Code or Codex subscription, or a local model. It operates the apps you allowed on a paired Android phone and asks before publishing.
