Your phone already holds your messages, your shopping, your travel plans, and your work. Mika is an open source Android app that turns a language model into an agent that operates those apps for you, on a hidden background screen, while you keep using your phone normally. You type one sentence, and it opens apps, reads the screen, taps, types, and scrolls until the task is done (wangzan101/Mika, GitHub, retrieved 2026-11-03). The project was created on October 29, 2026, is licensed under Apache 2.0, and is explicitly an early public beta, not a finished product. It is a phone agent you can inspect and run with a model you choose, with limits worth reading before you hand it your phone.
Key Takeaways
Mika is an open source, on-device AI agent for Android, created on October 29, 2026 and licensed under Apache 2.0.
It works on a hidden background screen, so it can operate apps and websites without taking over the screen you are using.
There is no Mika server. You bring your own model API key or sign in with a ChatGPT, xAI, or OpenRouter account, and your requests go to the provider you configure.
By default it runs in Ask mode, which checks with you before spending money. Autonomous mode removes even that prompt, so most people should stay in Ask mode.
It is an early public beta: Android 11 or later for chat, web, and files, Android 12 or later for operating apps, and Android 14 or later recommended.
Phone agents are not a new idea, and big labs already ship their own versions. What makes Mika worth a look is an open license, a bring-your-own-model design, and a hidden screen that solves the most annoying part of phone automation: normally, an agent takes over your phone while it works. Mika does not. It also arrives with unusual momentum for a project only days old. Independent coverage noted about 1,682 stars two days after creation (explainx.ai), and the repository showed 5,230 stars and 1,005 forks at retrieval on November 3, 2026 (wangzan101/Mika, GitHub, retrieved 2026-11-03). The speed of that rise is why this project, rather than the many phone agent demos before it, is worth examining.
What Is Mika?
Mika is an on-device AI agent for Android that operates apps and the web to get things done for you. That is how its repository describes it, and each word in that sentence carries weight. On-device means the app lives on your phone, not on a remote server run by the project. Operates apps means it does not call an official API for each service: it looks at the screen and finds buttons the way a person does, so it can work with any app or website, with no per-app integration. To get things done means the unit of work is a sentence from you, such as sending a message or finding a product and adding it to a cart.
The repository lists demos that show the intended scope. One demo sends a message on a background screen. Another summarizes what a group talked about today. A third finds the cheapest nonstop flight from San Francisco to New York. A fourth picks a charger under 30 dollars rated 4.5 stars or more and adds it to the cart while stopping before checkout. Several demos use apps popular in China, such as WeChat and Xiaohongshu, but the project states it is not tied to them. Its built-in knowledge also covers Gmail, WhatsApp, Telegram, Uber, Amazon, and YouTube, and you can talk to it in English.
Aspect | Mika |
|---|---|
Created | October 29, 2026 |
License | Apache 2.0 |
Language and UI | Kotlin with Jetpack Compose; app UI in English, Chinese, Japanese, Korean, Spanish, Portuguese, Russian, and Indonesian |
Stars and forks | 5,230 stars and 1,005 forks at retrieval on November 3, 2026 |
Android needed | 11 or later for chat, web, and files; 12 or later for operating apps; 14 or later recommended |
Model | Bring your own key, or sign in with ChatGPT, xAI, or OpenRouter; no Mika server |
Status | Early public beta |
How the Hidden Screen Works
The hidden screen is the core idea, and it is also why setup is unusual. Mika opens target apps on a separate virtual display with its own focus, while you keep using your main screen. A small live view inside the chat shows what it is doing. Because the background screen has its own focus, the agent does not get in the way of your typing, which is why Android 14 or later is recommended for operating apps.
Perception starts with the accessibility tree, which exposes buttons, text fields, and other controls in structured form. For interfaces it cannot read that way, such as custom-drawn controls or unlabeled buttons, it falls back to on-screen text recognition bundled in the APK, and with a vision-capable model it can locate buttons from screenshots. The loop is the same one desktop computer-use agents use: observe, act, check, and repeat, applied here to Android apps on a physical device.
Starting a helper that can launch apps on a second display needs elevated shell permissions, so Mika asks you to pair Android's own wireless debugging once, through a setup wizard in Settings. Android normally turns it off when you change Wi-Fi or restart, and with your permission Mika can turn it back on. You can also start the helper over USB. That permission is powerful, which is why the risks section below matters as much as the features.
What It Can Do
Beyond one-off tasks, Mika bundles several features that make it more than a demo:
Operate apps. Read long lists and chat histories, tap and type, and share text into apps. It hands logins, verification codes, and payments to you.
Browse the web. A built-in desktop-class browser keeps you signed in, reads pages, downloads and uploads, and hands anti-bot block pages to you instead of forcing through.
Scheduled and conditional tasks. Ask it to check the weather at 8 every morning and tell you only if it will rain, or to do a task every Friday.
Memory and skills. It remembers preferences you tell it, and turns finished tasks into skills so the next similar task is faster.
Files and computing. A Linux sandbox per conversation runs Python, Node, and shell commands on files you send. It is arm64 only and offline by default.
Every task leaves a step-by-step record, answers list their sources, and each task ends with a summary of time, model round trips, and tokens used. That matters for cost: a phone agent reads the screen at every step, so it makes many model calls per task. Independent coverage advises starting with a cheap model on a small task and checking the numbers before scheduling anything recurring (explainx.ai).
How to Set It Up
The repository gives three steps, and each one has a catch worth knowing in advance:
Install. Download the APK from the repository's Releases page, which the project names as its only update source. Do not install Mika from anywhere else.
Connect a model. In Settings, pick a provider and enter an API key, or sign in with a ChatGPT, xAI, or OpenRouter account. Presets cover DeepSeek, Qwen, Kimi, OpenRouter, Anthropic, Gemini, OpenAI, xAI, Mistral, Groq, Ollama, and LM Studio, and a custom option takes any compatible endpoint, including a model on your own network.
Set up phone control. Follow the wizard in Settings and pair wireless debugging once, or start the helper from a computer over USB. Skip this step if you only want chat, web, and files.
If you are unsure your phone is up to it, the app includes a 12-check compatibility self-check in Settings that runs in about a minute. Some phone systems leave wireless debugging out entirely, in which case USB is the fallback. Models need tool calling to work as the main model.
The MCP Bridge: Your Desktop Agent Can Drive Your Phone
The feature most relevant to readers who already use Claude Code or Codex is the computer bridge. With the phone plugged in, two commands expose an MCP endpoint on the local machine for your desktop AI tool to hand Mika tasks, check progress, and fetch results. The endpoint is USB-only and token-protected, and payments can only be approved on the phone. If MCP is new to you, our CVE MCP Server article shows the same protocol idea in a different job: giving an agent a tool it can call. Mika goes the other direction from phone-as-remote-control tools: the agent lives on the phone, and the computer delegates to it.
Risks Worth Reading Before You Install
A phone agent combines screen access, logged-in accounts, and the ability to send messages and spend money, so the defaults deserve scrutiny. By default Mika runs in Ask mode: it checks with you only before spending money, and just does everything else. Sending messages, posting, submitting forms, deleting, and changing settings happen without asking. Autonomous mode removes even the payment prompt. For scheduled tasks that run while you are away, Ask mode skips the money-spending step and reports it. Most people should stay in Ask mode.
The privacy claim is strong and narrow at the same time. There is no Mika server, keys and memories are stored encrypted on the phone, conversations stay on the device, and update checks only contact GitHub. But what you say and what is on your screen still go to the model provider you configure, so the real privacy question is that provider's policy. A model reading your messaging apps sees whatever is on screen, including other people's messages. A local model through Ollama or LM Studio keeps data on your network, at the cost of capability.
Two more warnings come straight from the project. Operating third-party apps automatically may break their terms of service, with a risk of restricted features or account bans, so do not point it at an account you cannot afford to lose. The project is an early public beta, mainly tested on one Android 14 phone. Independent coverage adds the practical version: use a spare phone or a secondary profile with no banking apps and no important accounts logged in while you evaluate it (explainx.ai).
How Mika Compares
Desktop computer-use tools show the contrast. Our Vercel agent-desktop article covers an open source tool that lets any AI agent operate a computer without screenshots. Mika is the phone-side equivalent: the same idea on a device you carry, with a hidden screen instead of a visible takeover. Neither replaces a first-party assistant from a phone maker, which can use private system hooks Mika cannot. The trade Mika offers is that it is inspectable, under Apache 2.0, and runs on the model you pick.
The honest summary is that Mika is a well-scoped prototype with a genuinely useful trick, not a mature assistant. The hidden screen, the bring-your-own-model design, and the MCP bridge are real choices. The beta label, the Ask mode defaults, and the terms-of-service risk set the terms for trying it: on a spare device, in Ask mode, with a cheap model, on tasks where a wrong tap is annoying rather than expensive.
Frequently Asked Questions
Is Mika free? The app is free and open source under Apache 2.0. You pay your model provider for the model calls, or use a ChatGPT Plus or Pro sign-in. There is no Mika server and no Mika subscription.
Which Android phones can run Mika? Chat, web, and files need Android 11 or later. Operating apps needs Android 12 or later, with Android 14 or later recommended. The Linux sandbox needs an arm64 device.
Does Mika send my data to its own server? No. The project states there is no Mika server: what you say and what is on your screen go only to the model provider you configure. Keys and memories are stored encrypted on the phone, and conversations stay on the device.
What happens at a payment or login screen? Mika hands those steps to you. Payment and password pages often block screenshots, so Mika asks you to take over there and finish the step yourself. In Ask mode it also checks with you before spending money.
Can my desktop coding agent use Mika? Yes, through the MCP bridge. With the phone plugged in over USB, a desktop AI tool can hand Mika tasks, check progress, and fetch results. The endpoint is token-protected and USB-only by default, and payments can only be approved on the phone.
The Bottom Line
Mika is the most interesting phone agent to appear in a crowded month because it removes the worst part of the experience: losing your phone to the agent while it works. The hidden screen, the open license, and the bring-your-own-model design make it easy to inspect and cheap to try. It is also, by its own description, an early beta that asks for powerful permissions and acts without asking except before spending money. Try it on a spare phone, in Ask mode, with a cheap model, starting small. Start with the GitHub repository and its Releases page for the APK, or read the Indonesian version of this article.