Description
You ask an AI to send a WeChat message or summarize today's group chat, and all you get back is advice. You are still the one tapping the buttons. Mika is an AI assistant for Android that takes a one-sentence request and opens the apps, reads the screen, taps and types until it is done.
It works on a hidden background screen, so you keep using your phone. A small live view in the chat shows what it is doing, and when a step needs your own hands, like a login, a captcha or a payment, it sends a notification, lets you take over, and picks up again after you hand it back.
Operate any app: No per-app setup. It looks at the screen like a person, finds buttons, and reads long lists and chat histories screen by screen. For apps whose UI can't be read, it falls back to bundled OCR, and with a vision model it can find buttons from the image.
Built-in browser: Opens any site in the background and drives it with real clicks and keystrokes, keeping your logins. It paces page changes and hands risk-control pages to you instead of pushing through.
Floating capsule: While you're in another app, a capsule at the screen edge shows progress; tap to chip in, approve or stop, or double-tap to send the current screen, for example to ask how to reply to a message.
Your model, your key: There is no Mika server. Requests go only to the provider you configure, with OpenAI-compatible, Anthropic and Gemini endpoints supported, or sign in with ChatGPT.
Local Linux sandbox: Scripts run in a sandbox on the phone, on arm64 devices only.
Use it from your computer: Connect over MCP so an AI tool on your computer can hand Mika tasks, check progress and fetch results, over USB with a token.
It is a public beta that needs Android 13 or newer, with Android 14+ recommended for phone control, and a one-time wireless debugging pairing in settings. The UI is in Simplified Chinese, Traditional Chinese and English, under Apache-2.0.
It works on a hidden background screen, so you keep using your phone. A small live view in the chat shows what it is doing, and when a step needs your own hands, like a login, a captcha or a payment, it sends a notification, lets you take over, and picks up again after you hand it back.
Features
Operate any app: No per-app setup. It looks at the screen like a person, finds buttons, and reads long lists and chat histories screen by screen. For apps whose UI can't be read, it falls back to bundled OCR, and with a vision model it can find buttons from the image.
Built-in browser: Opens any site in the background and drives it with real clicks and keystrokes, keeping your logins. It paces page changes and hands risk-control pages to you instead of pushing through.
Floating capsule: While you're in another app, a capsule at the screen edge shows progress; tap to chip in, approve or stop, or double-tap to send the current screen, for example to ask how to reply to a message.
Your model, your key: There is no Mika server. Requests go only to the provider you configure, with OpenAI-compatible, Anthropic and Gemini endpoints supported, or sign in with ChatGPT.
Local Linux sandbox: Scripts run in a sandbox on the phone, on arm64 devices only.
Use it from your computer: Connect over MCP so an AI tool on your computer can hand Mika tasks, check progress and fetch results, over USB with a token.
It is a public beta that needs Android 13 or newer, with Android 14+ recommended for phone control, and a one-time wireless debugging pairing in settings. The UI is in Simplified Chinese, Traditional Chinese and English, under Apache-2.0.
