v2.1 hotfix

• Commands are no longer executed twice.
• The database is no longer automatically transferred in its entirety; instead, the AI ​​can retrieve individual parts of it to prevent performance degradation caused by too many input tokens.

v2

New language models:
Mistral Large 3
GPT-5.4 Nano
GLM-5V Turbo
Gemini 3 Flash
Mistral Medium 3.1
Gemma 4 E4B it (offline) 3.4 GB
and more

• Offline models can be run on the GPU.

• A human expert can now be selected.

• The task now remains in the text field when
switching models.

• The task can now be easily repeated without having to re-enter it.

• API keys no longer need to be entered at startup.

• Top P, Top K, and temperature can now be set.

v1.2

Doesn't work with Android 14-
This will be fixed soon.

• Gemini 2.5 Flash live preview has been added. This significantly accelerates screen operation and is now the new standard.
• AI can now open apps instantly with a command.
• Responses from the language model are now streamed.
• No more toast when going back and pressing the home button.
• Incorrect recognition of words and sentences as commands has been fixed.
• The system message can now be restored.
• 503 error in Gemini 2.5 Pro has been fixed.
• Target SDK upgraded to Android 15 according to Google's specifications (for Play Store).
• Design improved, black bars removed.
• AI can now click and hold.
• Unnecessary information has been removed from screen elements.

Screen Operator v1.1.apk

Don't work with Android 13-

• Gemini 2.5 Flash API updated because Google doesn't automatically forward the API.
• New models added.
• Gemini 2.5 Pro now works without adding a billing account. Model overload issue with Gemini 2.5 Pro will be fixed soon.

Screen Operator v1.0.apk

First Github release