A 83MB "Doubao phone alternative": how Mika got 7000 stars in 8 days, and what room is left for those following

ByteDance and Nubia turned "AI taps your phone for you" into a 5999 yuan new phone. An indie developer on X, calling himself the "job-hunting edition", made the same thing into a free APK that runs on an old Android phone: the AI operates apps on an invisible background screen while you keep scrolling your own phone. Open-sourced 8 days ago, it has 7054 stars, 1594 forks, and roughly 46k downloads. Every component here has been built by someone before; what set it apart was the packaging and the launch-week distribution. Below is why it took off, and what room is left for anyone wanting a project that can earn stars just as well.

A tweet at 8 a.m.

On Oct 30, 2026, at 8 a.m. Beijing time, X account "王赞(求职版)" posted: "Mika is officially open source: a real Phone Agent." The demo shows scenarios every Chinese user gets at a glance: opening WeChat to send a message to File Transfer, summarizing a RedNote blogger's post, and batch-greeting recruiters on BOSS Zhipin.

This account was not new. It had already posted 1928 times before Mika, always mixing in Chinese AI circles. That day the author also posted an English version and later pinned it. About 11 hours after the launch, the English mega-account @RoundtableSpace restated it in English: "AI agents can already use computers, now there's one that can use your phone." Whether this was a paid promotion has not been verified.

That night the author reported roughly 400 stars and 1000 multiple downloads. By Nov 1 he said it had passed 3000 stars in under 60 hours, and his own X account had grown to 15k followers.

What it actually is

The first line of the README: "Your phone's AI assistant: tell it what you want, and it operates your apps and websites until the job is done." The second line holds the whole pitch: "It works on a hidden background screen, so you can keep using your phone."

Getting started takes three steps: download the APK from Releases; enter a model API Key, or sign in directly with a ChatGPT, xAI or OpenRouter account; and if you want it to operate other apps, pair wireless debugging on the phone once using the setup wizard. No root, no computer.

The mechanism was explained by @0xLogicrw in a long technical post: the Doubao phone assistant relies on system-level permissions granted by the phone maker. Mika instead uses Android wireless debugging to launch an auxiliary process with shell permissions, creating an invisible virtual screen on an ordinary phone where apps run. The interface is read through accessibility, with built-in text recognition as a fallback when that fails.

Around this core it stacks a whole set of features: a built-in desktop-style web browser, scheduled tasks ("check the weather every morning, only tell me if it's raining"), a floating capsule (double-tap to send the current screen and ask "how should I reply to this"), memory and skill distillation, a Linux sandbox for each conversation, MCP, and assigning it tasks from WeChat. It presets more than twenty model providers, including DeepSeek, Tongyi, Kimi and Zhipu, and can also connect to a self-hosted Ollama on the local network. There is no Mika server of its own. Apache-2.0. The README is in six languages.

Media installed it and tested it

On Nov 1, Smzdm published "Doubao phone alternative: an open-source app turns an Android phone into an AI phone." The author installed the 83MB APK on a Redmi K50 (Android 14), using a model self-hosted on the local network, and fought with it for a full day:

• "Set alarms for tomorrow morning at 9:40, 9:50 and 10:00": 7 steps, 1 min 52 sec, 126k tokens, and it succeeded. At step 4 opening the clock failed, so it switched on its own to looking for the entry on screen. The author checked the system's underlying data: all three alarms were there, and the original alarms were untouched.
• "Search JD for the price of the iPhone Duo": 21 steps, 8 min 23 sec, 1.18M tokens, also succeeded, and it proactively warned that the results mixed in other brands' foldables and used phones.
• Scheduled tasks also worked.

This test showed two things. First, it really works, and word of mouth spreads because media tests confirmed it. Second, it is slow and token-hungry: one price lookup took 8 minutes and millions of tokens. The most discussed GitHub issues say the same: "it uses a bit too many tokens", "add a step cap / token budget cap per task so long tasks don't drain the quota", and the local Ollama context balloons and times out.

How the stars climbed

(Daily GitHub star records; the 10-30 figures come from the author's tweets; download counts come from the Release download badge, 11-06.)

A few things stand out:

• Both language tracks burned at once. On the Chinese side: the author's own account, @0xLogicrw, @QingQ77, @indie_maker_fox and other Chinese AI-circle accounts, then Smzdm, CSDN and Bilibili. On the English side: @RoundtableSpace, AlphaSignal, various AI newsletters and aggregators, then the "comment ANDROID for the link" Reels on Instagram. Both sides took off within 48 hours.
• It didn't rely on GitHub's own traffic. A new repo under two months old can't get onto trending, so all of these 7000 stars came from outside the platform. No record of a Product Hunt or Hacker News front-page placement was found.
• It isn't just bookmarks. 46k downloads, roughly 6.5 downloads for every 1 stars; the fork-to-star ratio is about 23%, roughly double that of new repos of similar size. There are many derivatives, such as mylea1/phongent, created on 11-01, "a multi-device agent controller based on Mika", with 343 stars. The author's account has only this one repo, which looks like the "new account, single repo surge" pattern of star farming, but the downloads, forks and a pile of spontaneous reviews don't match star farming, so the judgment is real heat.
• The release cadence is itself content. In 8 days it shipped to v0.1.36, and the author posted a new-feature tweet almost every day: voice, MCP, WeChat integration, each one another round of exposure. The app also hides an easter egg: have Mika open WeChat and message Wang Zan, and you get into the beta group. It serves as both a demo and a way to build the group.

Competitors are no shortage; where it wins

"Let AI operate apps on the phone" already had a long list of open-source projects before it:

The last row is the most telling. Ruto-GLM launched 10 months before Mika with nearly the same pitch and got 735 stars. Each of Mika's three points (a background virtual screen, no root and no computer, any app with any model) has been done by someone before. What Mika did was be the first to pack all three into an APK an ordinary person can open and use, add a setup wizard, GIF demos and a six-language README, and then have someone with an audience ship it at the right moment.

The right moment came from the Doubao phone. The first generation came out in December 2025, and the new device went on sale in Sep 2026 16. "AI taps your phone for you" had already gone mainstream among Chinese users, but the entry price was a 6000 yuan new phone. The first generation also triggered risk controls from WeChat and banks over cross-app operations and pulled its financial-category operations offline. So the title "Doubao phone alternative" was ready-made, and Smzdm and @0xLogicrw both used it. On the English side, Google's I/O talk about agentic Android helped set the backdrop.

So differentiation is the entry ticket: without it you can't say "you scroll your phone, it does its thing." But what separated 735 from 7000 was execution and the distribution in those two launch days.

Open questions

• Of the 46k downloads, how many people actually went through the "enable developer options, pair wireless debugging" steps, and how many still use it, is not visible from outside. Android 14 and up gives a better experience, and that bar filters out a lot of people.
• Whether apps like WeChat will put users' accounts under risk control because of automated operation. The author's own disclaimer warns it may violate third-party terms.
• The account is called "job-hunting edition", so the project is partly a portfolio piece (inference). Whether the author will keep shipping this way every day after he finds a job is uncertain.

If it were you

One-line verdict: building another "Android Mika" makes no sense, since that spot is taken, with four or five hundred added per day. But the category has just been lit up, and there are still open patches beside it. The more worthwhile thing to copy is its playbook: take a new concept that big companies gate behind hardware, and ship a free open-source alternative that works right after install. It can be tried, provided you already have a distribution channel.

Who it suits: developers who already have a few thousand followers on X or Chinese tech communities; people who have done Android system-layer work (ADB, accessibility, virtual screens) or native HarmonyOS development; and people willing to ship and post every day for one to two weeks after launch.

Who it doesn't suit: solo developers with no distribution channel who hope to be seen through GitHub's organic traffic; Ruto-GLM is the example. It also doesn't suit anyone aiming to make money from it: an open-source phone agent is free by nature, and users still pay for their own model.

How to do it (pick one):

• Switch platform: pure HarmonyOS. HarmonyOS NEXT can't install APKs, and Huawei users are the ones most heavily primed by the "AI phone" pitch. A HarmonyOS-native, no-PC "tap apps for you" assistant doesn't seem to exist yet (the only thing I found was Huawei's official developer capability market). The catch is that HarmonyOS's permission model differs from Android's, so background operation has to be verified first.
• Do what it doesn't do well: cut token use, speed it up. Checking a price takes 8 minutes and 1.18M tokens, and that's the complaint users raise most. A design where a user records an action once, it's saved as a skill, and the model no longer re-reads the screen next time (someone already raised this in the issues), or a version that sets a budget per task and automatically switches to a cheaper model, can both be explained in one sentence. It can ship as a plugin or skill library for Mika to ride its traffic, or stand alone.
• Build its ecosystem: an "awesome-mika"-style library of app-operation knowledge and skill collections. Mika lowered contribution to plain text, so whoever first organizes these into a library can capture spillover from its 1600 forks (inference).
• Apply the playbook to the next big-company concept: whenever a big company launches an AI feature that requires buying hardware or only works inside its own ecosystem, start building an "open-source, free alternative" the same day.

How to get users: follow Mika's cadence. At launch, post a one-sentence explanation of the technical trick, with a localized GIF (WeChat and food delivery for Chinese, Amazon and Google Flights for English). Post Chinese and English versions the same day, and contact an English AI news influencer in advance. Title it directly "XX alternative". Give verified-working install packages to places like Smzdm, Sspai, and CSDN, where people actually install and test. After launch, ship a new version daily with one post per version. Put an Easter egg in the app, like "let it add you to the group chat itself," that both demos the product and pulls people into the group.

When to launch, before what: Mika's heat is still going, with four to five hundred new stars a day. Ecosystem products (skill libraries, plugins) are best shipped within a month, i.e., before early December. The HarmonyOS and token-saving versions are less bound to Mika's momentum, but there's probably a quarter or two before big companies make the same capability free in their systems (inference).

What would show I'm wrong: ① A month from now, Mika's daily stars drop to a few dozen, and the issues are mostly "can't pair" and "installed but don't know how to use it": that would mean consumer phone-agent interest is mostly spectator interest, and followers won't keep earning stars. ② Huawei or Xiaomi builds "background operation of any app" straight into the system and opens it to older models: the alternative rationale disappears. ③ Major apps like WeChat start banning automation accounts in bulk: the category as a whole cools off.

Data and sources

GitHub repo page, README, and daily star history (10-31 to 11-06); Release download badge (11-06); X: @wangzan101 10-30 launch tweet and follow-ups, @RoundtableSpace 10-30, @0xLogicrw 10-31, @indie_maker_fox; Smzdm, "Doubao Phone Alternative" (11-01); AlphaSignal coverage (via hermes-ai.net repost), Upcurrent newsletter (10-31); CSDN source code walkthrough; YouTube re-uploads; GitHub Issues (11-06). Competitor star counts as of 11-06; "about" means rounded.