Mika Review: A Free Android AI Agent That Uses Your Phone for You

Your phone is the one computer an AI agent cannot reach. Coding agents live in your terminal, desktop agents drive your browser, and the phone in your pocket stays a manual device. Mika is the latest open-source project to attack that gap head-on: an AI assistant that runs on your Android phone and operates your real apps on a hidden background screen while you keep using the phone yourself.

Created on October 29, 2026, it passed 5,300 GitHub stars within five days. That kind of velocity usually means one of two things: a genuine hole in the market, or a demo that flatters more than it delivers. This review digs into which one Mika is, what it costs to run, and the catches hiding in the setup wizard.

What Mika actually is

Mika describes itself in one line: "An AI assistant on your phone: say what you need, and it works your apps and the web until it's done." You type a plain-language request, and instead of answering with text, it opens apps, taps buttons, types into fields, and scrolls pages, on a separate background screen that Android carves out as a virtual display. Your own screen stays yours. A small live view lets you peek at what the agent is doing whenever you want.

The demo reel shows the intended pitch: "Send a message to File Transfer" opens WeChat on the background screen, finds the contact, and sends the message. "What did the group talk about today?" scrolls through a group chat and summarizes it. "A charger under $30, rated 4.5 stars or more" picks a charger on a shopping site, adds it to the cart, and stops before checkout. The ordering takeout demo stacks discounts and asks before paying. The honest read of the reel: the agent does the reading and navigating, and hands the money moments back to you.

Beyond apps, Mika ships a built-in desktop-class browser that keeps you signed in, a per-conversation Linux sandbox that runs Python, Node, and shell on arm64 devices, and scheduled tasks ("check the weather at 8 every morning and tell me only if it will rain"). It can also be driven from a computer: AI tools on your desktop hand Mika tasks through a CLI, an agent skill, or MCP.

Why it blew up in five days

Verified live on November 3 via the GitHub API: 5,302 stars, 1,028 forks, Apache-2.0 license, first commit October 29, 2026, pushed November 2. The project is written in Kotlin with Jetpack Compose and tagged for accessibility, AI agents, Android automation, and mobile agents.

The traction is not just a counter. SteadyVane covered it in its October 31 daily brief, AlphaSignal ran an "invisibly in the background" piece the same day, and independent writeups followed: an explainx.ai setup and risk guide, a buildrun.io setup walkthrough, an aiidelist guide to its MCP bridge, and a November 1 hands-on blog from Yi. That is a lot of independent coverage for a project that did not exist the previous week.

The timing explains the appetite. Phone control has been the missing piece in the agent wave: OpenClaw lets you text an agent from WhatsApp but it lives on your own computer, and our Ghosthand vs Glasshand piece covered three ways to hand an agent an iPhone, two of which cost $60 to $99 a month for a hosted device. Mika is the Android answer, and it costs nothing for the app itself.

What it costs: free app, your model bill

Mika is free under the Apache-2.0 license. The APK comes from the project's own GitHub Releases page, there is no subscription, no credit pack, no paid tier. The pricing story has one layer that matters: Mika has no server and no model of its own. You bring your own model provider, and that provider bills you per token.

The setup screen offers presets for DeepSeek, Kimi, MiniMax, Qwen, Zhipu, OpenRouter, Anthropic, Gemini, OpenAI, xAI, Mistral, Groq, Ollama, and LM Studio, plus a Custom option for any OpenAI-compatible, Anthropic, Gemini, or OpenAI Responses endpoint. You can also skip the API key and sign in with a ChatGPT Plus or Pro account, an xAI/Grok account, or an OpenRouter account.

So what does a task actually cost? Take DeepSeek V4.1 Flash, a cheap mainstream option at off-peak rates of $0.15 per million input tokens and $0.60 per million output tokens (verified against DeepSeek's official API docs in our DeepSeek V4.1 Flash review). A routine phone task of roughly 10,000 input and 2,000 output tokens works out to about $0.003, or a third of a cent. In pounds and loonies at November 3, 2026 market rates (GBP 0.746, CAD 1.398, verified at publication), that is roughly 0.2p and C$0.004. The honest fine print: this is my arithmetic from two verified datapoints, not a published per-task price, and it swings with the model you pick. A frontier model doing vision-heavy app operation can cost 10 to 30 times more per task.

The pricing table, in short:

There is no world in which Mika itself charges you. The thing to budget is tokens, and the dial for that is which model you connect.

Ask mode vs Autonomous mode: the safety valve

By default Mika runs in Ask mode: it checks with you only before spending money (placing orders, paying) and otherwise just does the work. You can switch to Autonomous mode, which removes the checkpoints. For anything involving money, logins, or other people's inboxes, Ask mode is the sane default, and the demos are built around it: the takeout demo "orders takeout with stacked discounts, asks before paying," and the charger demo adds to the cart and stops before checkout.

Two hard boundaries come from the README itself, and they are worth quoting. First: "Operating third-party apps and websites automatically may break their terms of service, so use your own judgment." Automating someone else's app is your legal exposure, not Mika's. Second: screens that block screenshots, like payments and password entry, cannot be read by the agent, so Mika hands those moments to you by design.

Setup: three steps and one pairing

1. Install: download the APK from the project's GitHub Releases page. There is no Play Store listing; updates come from the repo's Releases or the app's own About screen.

2. Connect a model: in Settings, pick a provider preset and enter an API key, or sign in with a ChatGPT, xAI, or OpenRouter account.

3. Enable phone control: run the "Set up phone control" wizard and pair wireless debugging on the phone once. Skip this if you only want chat, web, and files.

The requirements table, straight from the README:

Unsure your phone qualifies? Settings includes a compatibility self-check: 12 checks in about a minute, with a one-tap report you can send the project.

The known annoyance, flagged in the project's own FAQ: Android turns wireless debugging off when you change Wi-Fi networks or restart the phone. With your permission, Mika turns it back on by itself; if it cannot connect, you go back and pair again. Expect this to be a recurring chore on phones that roam between networks.

Privacy: the honest version of "on-device"

Mika's README says "your data stays on your phone" and "there is no Mika server." Both are true with a caveat that matters: what you say and what is on your screen go to the model provider you configured. There is no Mika server in the middle, but "on-device" does not mean local inference. If your provider is a cloud API, your screen contents travel to that cloud.

The rest of the privacy story is good: API keys and memories are stored encrypted on the phone, conversations and run records stay on the device, and the only network call the app makes on its own is checking GitHub for updates. Treat the app as a well-behaved client of your model provider, not as a sealed box.

What it cannot do yet

The built-in app and website knowledge is written in Chinese, and several demos use apps popular in China (WeChat, Xiaohongshu). The README says this does not limit the agent, and the app UI ships in English, Spanish, Portuguese, Russian, Indonesian, Japanese, Korean, and both Chinese variants. But a first-week project whose reference demos lean Chinese is a fair caveat for Western-app reliability: English-app behavior has less public evidence behind it.

There are no published task-success benchmarks, only demos. The project's own FAQ lists the failure modes it knows about: payment and password screens it cannot read, wireless debugging that drops, and third-party app updates that can break its screen-reading. Star counts measure attention, not reliability, and five days of attention is not a track record.

How it compares

Mika's edge is the combination: it is the only one of the four that puts the agent on the phone you already own, at zero software cost, with no per-month seat for the device itself.

Pro tips

Start and stay in Ask mode: the one setting that prevents a 3 a.m. surprise purchase is the default, but check it is still the default after updates.

Run the compatibility self-check before you get attached: the 12-check report takes a minute and tells you whether app operation will even work on your phone.

Put your cheapest model on routine tasks: summaries and scheduling do not need a frontier model. Save the expensive provider for vision-heavy app operation where the cheap one fails.

Pair on your most stable Wi-Fi: wireless debugging drops on network changes and reboots, so do the one-time pairing where your phone spends most of its time.

Read an app's ToS before automating it: Mika's own disclaimer puts this on you. Anything with an account, a wallet, or other people's data deserves the Ask-mode checkpoint.

Hand phone jobs from your desktop agent via MCP: the CLI and MCP bridge exists so a coding agent on your computer can delegate "do this on my phone" instead of you typing it twice.

FAQ

Is Mika really free?

Yes, the app is free and open source under the Apache-2.0 license, with no subscription or paid tier. Your costs are whatever your chosen model provider charges per token, which is separate from Mika entirely.

What phones does Mika work on?

Chat, web, and file features need Android 11 or later. Operating other apps needs Android 12 or later, with Android 14 or later recommended so the background screen does not fight your own typing. The Linux sandbox needs an arm64 device.

Do I need a paid AI subscription?

Not necessarily. You can bring a cheap API key (DeepSeek, Kimi, MiniMax, OpenRouter presets exist), point it at a local model via Ollama or LM Studio, or sign in with an existing ChatGPT Plus/Pro, Grok, or OpenRouter account. There is no Mika-side subscription either way.

Is my data sent to Mika's servers?

There are no Mika servers. Your words and screen contents go to the model provider you configured, so if that provider is a cloud API, your data leaves the phone to that provider. Keys and memories are encrypted on the device, and conversations stay on it.

Can Mika pay for things or log in for me?

Payment and password screens block screenshots, so the agent cannot read them and hands those moments to you. In Ask mode it also stops before spending money. That is a feature, not a bug.

How is Mika different from OpenClaw?

OpenClaw is a personal assistant runtime that lives on your computer and takes orders over WhatsApp or Telegram. Mika lives on your Android phone and operates the phone's own apps on a hidden background screen. Different homes, different jobs.

The takeaway

Mika is the most credible attempt yet at the phone-shaped hole in the agent wave: a free, open-source Android agent that runs your apps while you keep using the phone. The five-day star rush is real attention on a real gap, but attention is not reliability, and the honest price is your model provider's token bill plus a wireless-debugging pairing you will occasionally redo. If you have an Android 14 phone and a cheap API key, this is the weekend project worth trying. This is general information, not legal advice: automating third-party apps can breach their terms of service, and that judgment call stays with you.

Sources

• wangzan101/Mika on GitHub: official repository and README (hidden background screen, Ask/Autonomous modes, provider presets, requirements, privacy and ToS statements; star and fork counts verified live via the GitHub API, November 3, 2026)
• Mika Releases: APK distribution and version history
• explainx.ai: setup walkthrough and risk notes
• haoyi.im (November 1, 2026): independent hands-on walkthrough
• SteadyVane (October 31, 2026): daily brief coverage
• AlphaSignal (October 2026): launch coverage
• buildrun.io: setup guide
• aiidelist: MCP bridge and privacy guide
• AI Core Updates: DeepSeek V4.1 Flash 1017 review: DeepSeek off-peak rates $0.15/$0.60 per million tokens, verified against DeepSeek's official API docs October 23, 2026