Mika Turns Your Android Phone Into an AI Agent. Check What It Does Without Asking.
An open-source app that works your apps on a hidden screen, with your own model key. Its safety default guards your wallet and nothing else.
A phone agent that can send a WhatsApp message for you is useful exactly once. The second time, you start wondering what else it will send.
That is the real question behind every AI agent that operates apps on your behalf, and most of them answer it in a settings screen nobody reads. The model gets the attention. The confirmation rules decide what happens when the model is wrong.
Mika is a good place to learn this, because it answers the question unusually plainly. It's an open-source Android app, trending on GitHub this week, that takes a one-sentence request and carries it out by tapping through your apps on a hidden background screen while you keep using the phone. It shipped ten releases on October 30 alone, v0.1.3 through v0.1.12. And its README states the rule most agents bury: by default, only the step that spends money asks you first. Sending messages, posting, submitting forms, deleting things and changing settings all go through without a prompt.
That isn't a scandal. It's a design choice, written down honestly. But it's the wrong default for most people, and once you see why, you'll know how to evaluate every phone agent that comes after it.
What Mika actually is
In plain terms: Mika is an app you sideload (install from a downloaded file, outside the Play Store) that turns a sentence like "order my usual from Uber Eats" or "reply to Sam on Telegram that I'm running late" into a series of taps and typing.
A few things make it different from the phone assistants you already have.
It doesn't bundle a chat model. You plug in your own model through an API key (a password that lets an app use a paid AI service on your account). The README lists presets for DeepSeek, Qwen, Kimi, Zhipu, OpenRouter, Anthropic, Gemini, OpenAI, Ollama and LM Studio, and any compatible endpoint (the web address an AI service answers on) works. A ChatGPT Plus or Pro login is an alternative. The model has to support tool calling, which means it can ask the app to take actions instead of only writing text. (The app does ship on-device text recognition for reading screens, but the thinking happens in the model you pick.)
There's no Mika server. Per the README, your messages, screen content and screenshots go only to the model provider you chose. Your API keys, memory and user profile are encrypted with the Android keystore (the phone's built-in locked storage for secrets), and chat history and logs stay on the device.
It works on a screen you can't see. Mika runs tasks on a hidden background display, so it isn't fighting you for the phone.
It has a small Linux computer inside. Each conversation gets a sandbox (a fenced-off workspace) that can run Python, Node and shell commands on files you send it. That sandbox is offline by default; scripts only reach the internet if you switch on "allow scripts to use the network."
The license is Apache-2.0. The latest release, v0.1.12 from October 30, adds built-in notes for apps like Uber, Uber Eats, WhatsApp, Telegram and Chrome, and for sites like Amazon, Gmail, Google Calendar and Google Maps, so tasks there run faster and fail less (release notes).
How it gets control of your phone
This is the part to understand before you install it, because it explains how much power you're handing over.
Android doesn't let one app tap buttons inside another app. To get around that, Mika uses two system features.
The first is Accessibility, a permission built for screen readers that lets an app read what's on screen and act on it. Mika uses it to read the screen and tap.
The second is wireless debugging, a developer feature that normally lets a programmer's computer control a phone over Wi-Fi. Mika's setup guide walks you through turning on Developer options, keeping USB debugging on, enabling wireless debugging, and pairing with a six-digit code, all on the phone with no computer. Mika uses that connection to launch its built-in helper, which drives apps on the hidden screen.
The README doesn't spell out what privilege level that helper runs with. For context, Android's own developer docs describe the same pairing flow as the way a workstation gets debug access, including a shell that can start apps and run package-manager commands. Two details from those docs matter here: ticking "always allow on this network" makes that Wi-Fi a trusted debugging network, and a paired device stays paired until you remove it.
Mika's guide also suggests turning on "Disable adb authorization timeout" (adb is Android Debug Bridge, the tool behind this debug connection) so the pairing doesn't expire. That's convenient. It also means the access doesn't expire either.
None of this is hidden. It's simply a lot of power for a beta.
The two modes, and the one line that matters
According to the README, Mika has two modes. The switch lives under Developer options, labeled "autonomous mode": off means ask, on means autonomous.
Ask mode (the default). In the README's words, translated from Chinese: "only the step that spends money asks you first." Sending messages, posting, submitting forms, deleting and changing settings run without asking.
Autonomous mode. Payments also go through without asking. The README: "it doesn't ask before payment either, and just finishes."
In both modes, logins, verification codes and payment passwords always hand back to you. Payment and password screens often block screenshots, so Mika couldn't see them anyway. Scheduled tasks in ask mode skip any spending step and report it instead of waiting.
So here is the honest summary. Mika's default guards your money. It does not guard your reputation, your inbox, your contacts or your data.
The insight: map the irreversible actions, not the model
When people evaluate a phone agent, they ask how smart the model is. That's the wrong first question.
Models misread screens. They tap the wrong contact with the same name, paste a draft into the wrong chat, or "clean up" a folder a little too thoroughly. Mika's own README admits it's mainly tested on one Android 14 phone and that its demo clips are sped up 1.5 to 4 times. That isn't a knock on Mika. It's a reminder that every agent fails sometimes, so the useful question is what happens when it does.
Sort an agent's possible actions into two piles:
- Undoable: opening an app, searching, reading a screen, drafting text, adding to a cart.
- Irreversible or embarrassing: sending a message, posting publicly, submitting a form, deleting, paying, changing a setting you won't notice.
A good confirmation rule stops at the second pile. Mika's default stops at one item in it, payments. Everything else in that pile happens on the model's judgment.
This pattern isn't special to phones. A research benchmark for command-line agents released this month, RealCLI, found the best configuration averaged 49.2% on tasks and peaked at 37.6% on working together with a person, with models weakest at steerability, trust calibration and safety boundaries. Those numbers measure command-line agents, not Mika. But the pattern carries over: today's agents are worst at exactly the moment Mika's default skips, deciding when to stop and check with you.
Put this into practice
If you want to try Mika, here's a setup that gets you the useful parts with a smaller blast radius.
1. Use a spare phone first
This is the single best safety step. An old Android 13 or newer phone (Android 14 recommended; tasks need Android 13+, and the Linux sandbox needs an arm64 device, the 64-bit ARM chip type) with a fresh Google account and only the apps you want Mika to touch. Your real WhatsApp, your bank and your work email stay on your real phone.
2. Start without phone control at all
The README says pairing can be skipped if you only need web, files and chat. Do that first. Download the latest APK (the Android install file) from Releases, add a model provider, and use the chat and sandbox for a day. You'll see how your chosen model handles instructions before you give it hands.
3. Pick a cheap, capable model and watch the bill
Every screen Mika reads is sent to your model provider, so long tasks can use a lot of tokens (the units AI providers bill by). Mika keeps a local record of token usage. Check it after your first few tasks before you set anything to run on a schedule.
4. Keep ask mode, and phrase tasks so the risky step is yours
Since ask mode won't stop before a send, put the stop in the request itself. Instead of "reply to Sam that I'm late," try "draft a reply to Sam saying I'm late and leave it unsent." Instead of "clear out old screenshots," try "list screenshots older than six months." You get the agent's speed on the boring part and keep your thumb on the irreversible one. This depends on the model following your instruction, so check the first few results by hand.
5. Know how to take the keys back
From the setup guide: Settings, Phone operation, Unpair inside Mika, or remove Mika under the system's Wireless debugging, Paired devices. Android's docs also offer "Revoke adb debugging authorizations," which removes every paired device at once. Turn off wireless debugging when you're done testing, and leave "Disable adb authorization timeout" off on a phone you care about.
Honest limitations
Mika is a public beta. The README warns that features change quickly and upgrades may break compatibility, and it shipped ten versions in one day, which is either great momentum or a sign to wait a week, depending on your patience.
It's tested mainly on one phone. Your phone's maker may draw screens differently, and the agent can get lost.
It can't see screens that block screenshots, like payment and password fields, so some tasks will stall and hand back to you.
Automating other companies' apps may break their terms of service. The README says so directly, and warns that accounts can get restricted or banned. You carry that risk, not Mika.
And the confirmation gap is the big one. Until Mika lets you choose which action types require approval, the safety boundary is your wording, your spare phone and your choice of model.
The question to ask any agent before you hand it your phone
Mika deserves credit for writing its rules down where you can read them. An agent that never tells you its rule is harder to trust, not easier.
So use it as a template. Before you give any agent your phone, your inbox or your browser, find the line in its docs that says what it does without asking. If you can't find that line, assume the answer is "everything." If you can, decide whether the things on that list are things you'd let a new intern do unsupervised on day one. Then set the agent up to match your answer, not its default.
Sources: Mika README and setup guide (main branch, reviewed at v0.1.12), Mika releases, Android Debug Bridge docs, RealCLI on arXiv. Details as of October 30, 2026.