Mika is an on-device AI agent for Android, published as an open-source Kotlin application under Apache-2.0, that lets a phone user state a task in one sentence and then operates apps and the web on a hidden background screen to carry it out.

What it is

Mika is an Android application written in Kotlin and built on Jetpack Compose, according to its repository topics, and it is distributed as an APK through GitHub Releases rather than through a server component. There is no Mika server: the app runs on the phone, and it connects to a language model either through the user's own API key or by signing in with a ChatGPT, xAI or OpenRouter account. The agent looks at the screen and finds buttons the way a person does, so it works with any app or website on the device without app APIs and without per-app integration.

The concrete problem it solves is the manual work of driving a phone. Rather than tapping through a long chat history to summarise what a group discussed, scrolling a shopping app for a charger under $30 rated 4.5 stars or more, or opening a news site to read the top five stories, the user describes the outcome and Mika performs the taps, typing and scrolling itself. It sits in the Android automation space alongside the accessibility services that make screen reading possible, and it replaces the combination of per-app integrations and hand-driven repetition that such tasks otherwise require. The repository carries 6,173 stars and 1,200 forks with 114 open issues.

Key capabilities

• Runs tasks on a separate background screen while the phone stays in normal use, with a small live view that can be tapped to watch the work in progress.
• Operates any installed app without per-app integration, reading the screen and locating controls visually.
• Provides a built-in desktop-class browser that keeps the user signed in, reads pages, downloads and uploads files, paces itself, and hands block pages back to the user.
• Supports schedules and conditions, such as "Remind me at 9 tomorrow" or "do this every Friday", which run on time inside the original conversation.
• Turns finished tasks into reusable skills and remembers stated preferences, so repeat runs are faster.
• Accepts input through hold-to-talk on the chat input or a long-press on the floating capsule, using the phone's built-in speech recognition or a provider's speech model if the phone has none.
• Allows intervention mid-task: the user can keep talking to add or correct instructions, long-press to withdraw or edit a message Mika has not yet picked up, and say "continue" after a stop to resume from the reported point.

Who uses it and how

• Individuals automating everyday phone chores, such as setting two alarms in one request or checking the weather at 8 every morning and reporting only when rain is expected.
• People working inside messaging and social apps, with the project's own demos covering chat-history summaries and a creator's recent posts; the demos use WeChat and Xiaohongshu but the agent is not tied to them.
• Shoppers and travellers performing multi-step tasks with a stop condition, such as adding an item to a cart and stopping before checkout, or finding the cheapest nonstop from San Francisco to New York.
• Job seekers running outreach at small scale, as in the demo that greets recruiters for three jobs on a job site.
• Users who need human oversight at sensitive steps: logins, verification codes and payments are handed back to the person rather than completed by the agent.

Getting started

Download the APK from the latest GitHub Releases entry, then connect a model under Settings › Model by entering an API key or signing in with an account. To operate phone apps, follow the "Set up phone control" wizard in Settings and pair wireless debugging on the phone once; that step can be skipped if only chat, web and file use is wanted.

How it compares

No comparable tool is named in the facts supplied for this entry, and no list of paid products is given, so Mika stands alone in this registry on the evidence available. The only external services it is documented against are model providers — ChatGPT, xAI and OpenRouter — which supply inference rather than competing with the agent itself.

When to use it — and when not

Mika suits Android users who are willing to supply their own model credentials and to complete a one-time wireless debugging pairing, since both are prerequisites for the agent's core function. It is a poor fit for anyone who needs a fully unattended flow, because authentication, verification codes and payment steps are deliberately handed back to the person, and for anyone outside Android. The licence is clear and the repository is active, but the project also carries 114 open issues, which is worth reviewing before committing to it.