wangzan101/Mika

Mika is a public-beta Android app that turns a plain-language request into actions on your phone: it opens apps, reads screens, taps and types on a separate background screen, browses the web in a built-in desktop-mode browser, and can run scripts in a local Linux sandbox. It is aimed at people who want tasks like summarizing a chat group, checking a blogger's posts, or setting alarms done for them without losing use of the phone.

How the automation works

Mika operates any installed app without per-app integration by looking at the screen and finding buttons the way a person would, and it reads long lists and chat histories screen by screen. Apps that draw their own unnamed controls are handled with bundled text recognition, and an optional vision model can locate buttons visually. Work happens on a hidden background screen while a small floating capsule shows progress; you can interject, answer, approve or stop mid-task, and after stopping Mika reports what it did and where it left off so "continue" works later. Screen-reading and tapping rely on the accessibility service, which Mika tries to enable itself after setup.

Setup and prerequisites

Installing requires Android 11+, but running tasks requires Android 13+ (the background screen and built-in phone assistant), with Android 14+ recommended because the background screen then has its own focus. Web, sandbox and chat use are said to work on Android 11 and 12. To launch apps in the background, Mika borrows Android's wireless debugging: a one-time pairing in the setup guide, done entirely on the phone without a computer or cable; if wireless debugging turns off after a reboot or Wi-Fi change, Mika can reopen it with your permission. The Linux sandbox is arm64-only; other devices install but show it as unsupported. Web, file and chat use can skip the pairing step. Future Android versions or other phones are not validated: the author says it has mainly been tested on one Android 14 phone.

Models, privacy and approval

You bring a model API key. Presets cover several providers, and custom OpenAI-compatible, Anthropic or Gemini endpoints work, including LAN addresses such as http:// endpoints; an unencrypted non-LAN address triggers a warning about eavesdropping. Alternatively you can log in with a ChatGPT Plus or Pro account and use its quota. The main model must support tool calling; vision and lightweight models are optional. There is no Mika server: requests, screen content and screenshots go only to the configured provider, while keys, memory and usage history stay encrypted on the device. The model has only a sample of provider capability tags, marked unconfirmed when unspecified.

The default ask mode only pauses before spending money (ordering, payment), and payment itself is completed by you; messages, posts, submissions and deletions proceed without asking. Autonomous mode also skips the payment prompt. Logins, verification codes and payment passwords always require handoff in either mode, and scheduled tasks that would spend money simply skip that step rather than wait.

Risks and licence

The author warns that automating third-party apps and websites may violate their terms and lead to restrictions or account bans, and that you are responsible for actions taken on your behalf; check results for anything important. Pages that block screenshots (payment, password entry) cannot be seen by Mika and require takeover. The app changes quickly and upgrades may be incompatible. It is Apache-2.0, with bundled third-party components and sandbox runtime under their own licences.

Project source and documentation