Mika
Wang Zan (wangzan101) · Mika: Open-Source Android AI Agent That Operates Your Apps and the Web on a Hidden Background Screen
Mika is a free, open-source (Apache-2.0) Android app that takes a spoken or typed request and carries it out by operating your installed apps and websites on a hidden background screen, using a cloud or local model you connect yourself. It suits Android users comfortable with sideloading and developer settings who want a phone agent with no vendor server in the middle.
Best for
Technically comfortable Android 14+ users who want a phone agent that operates their own signed-in apps and websites, prefer an open-source app with no vendor server, and are willing to sideload an APK, enable developer settings and supply their own model
Not ideal for
iPhone users, people who want an app-store install that works without developer settings, anyone who needs confirmations before messages are sent or data is deleted, and users or employers that cannot accept beta software or possible terms-of-service issues on their main phone accounts
Who it's for
Android users, especially early adopters and developers, who want an on-device agent to operate their apps and the web with a model they choose
Capabilities
- Operates any installed Android app on a separate background screen, reading the screen and tapping, typing and scrolling while you keep using the main screen, with a live view you can open or close
- No per-app API integration: it finds buttons from the screen's accessibility data, bundled on-device text recognition for custom-drawn UIs, and screenshots when a vision-capable model is set
- Built-in desktop-mode browser that keeps your sign-ins, reads and operates pages, downloads and uploads files, paces requests and hands anti-bot block pages to you; an 'Ask Mika' bar works on the page you are viewing
- Ask mode (default) confirms orders and payments with a card showing where, what and how much; Autonomous mode skips that; in both modes logins, verification codes and payment passwords are handed to you
- Scheduled and conditional tasks (one-off, interval, daily, weekly, with conditions, cooldowns and idle-only runs) that report back in the original conversation, created by chat or by hand
- Floating capsule: double-tap to send the current screen to Mika ('how should I reply?'), long-press to speak, and see progress or approvals from inside other apps
- Bring your own model: presets for DeepSeek, Qwen, Kimi, Zhipu, MiniMax, OpenRouter, Anthropic, Gemini, OpenAI, xAI, Mistral, Groq, Ollama, LM Studio and more, plus any OpenAI-compatible, Anthropic, Gemini or OpenAI Responses endpoint
- Sign in with a ChatGPT Plus/Pro, xAI (SuperGrok or X Premium+) or OpenRouter account instead of pasting an API key
- Separate main, vision, light and speech-to-text model roles, with per-model context windows, compaction and reasoning effort
- Per-conversation Linux sandbox (arm64) that runs Python, Node and shell commands on files you send, offline unless you allow network access
- Explicit memory of facts and preferences you state, stored encrypted on the phone and editable, plus skills built from finished tasks so repeat jobs go faster
- Computer access: a CLI and agent Skill or an MCP server let AI tools on your computer send tasks, follow progress and answer Mika's questions over USB or a trusted Wi-Fi or private VPN, plus a browser task desk (`mika ui`)
- Connects to external MCP servers (remote, or local in the sandbox) with account authorization and tool allowlists
- App access list to keep Mika out of apps such as banking or payments, an optional step limit per task, and a step-by-step record of every run with token usage
Limitations
- Early public beta: the project's disclaimer says features and UI change quickly, upgrades may break things, and it has mainly been tested on one Android 14 phone
- Android only, installed as an APK from GitHub releases; no Play Store or iPhone version is documented
- Operating apps needs Android 12 or later (Android 11 covers only chat, web and files), and on Android 12 and 13 background work can interfere with focus on the main screen; Android 14 or later is recommended
- Phone control requires Developer options with USB debugging kept on and a one-time wireless-debugging pairing, or an adb command from a computer over USB (repeated after each restart if wireless debugging is unavailable), or root
- Ask mode only confirms steps that spend money; sending messages, posting, submitting forms, deleting and changing settings happen without a confirmation
- There is no bundled model: you need a tool-calling model through your own API key, a local server, or a supported account sign-in, and model usage is billed by that provider
- Screen contents and what you say are sent to the model provider you configure, and Mika's accessibility service and adb-level helper can read and act in any app it opens unless you exclude that app in the app access list
- The project warns that automating third-party apps and websites may violate their terms of service and risk restricted features or account bans
- Screens that block screenshots, such as payment and password entry, cannot be read, so you have to take over
- The built-in app and website knowledge is written in Chinese and weighted toward apps popular in China, though any app can be operated and the UI is available in 10 languages
- The Linux sandbox is limited to arm64 devices
- Computer access over Wi-Fi or a VPN uses plain HTTP with a token and should be enabled only on trusted networks
- Led by an individual developer with community contributors; no company, support agreement or published security audit stands behind it
Use cases
- Asking Mika to read through today's messages in a busy group chat and summarize what was discussed
- Having it shop for a product within a price and rating limit and add the pick to the cart, stopping before checkout
- Setting up a daily or conditional check, such as weather at 8 a.m. or a price drop, that reports back only when it matters
- Double-tapping the floating capsule on an open conversation and asking how to reply
- Comparing flight prices on a travel site in the built-in browser while you keep using the phone
- Letting a coding agent on your computer hand a phone-only task to Mika over MCP or the CLI and wait for the result
Our take
Mika is one of the more complete attempts at a general phone agent that actually runs on the phone. The background screen is the key design choice: tasks run on a virtual display, so you are not locked out of your own device while it works. Bring-your-own-model with no vendor server is a real privacy and cost advantage for people who already have API keys or a local model. The price is setup and trust. Phone control means sideloading, enabling USB debugging, pairing wireless debugging and handing an accessibility-level helper to software from an individual developer, first released as a beta in late October 2026. Ask mode guards money but not messages or deletions. Treat it as an enthusiast tool: try it on low-stakes apps or a spare Android 14 phone, use the app access list to fence off banking and work apps, and check results before trusting it with accounts that matter.
Who should use it
Android enthusiasts and developers on Android 14 or later who want an open-source agent to handle chores in their own apps and websites, people who already pay for model APIs or run Ollama or LM Studio and want to reuse them, and builders who want their computer-side agents to reach phone-only apps through MCP or a CLI.
Who should skip it
iPhone users, people who will not sideload apps or enable developer settings, anyone who needs confirmations before messages are sent or data is deleted, organizations that require vetted or supported software on work devices, and users whose important accounts could be harmed by terms-of-service enforcement against automation.
Strengths
- Free and Apache-2.0 open source, with no Mika server between your phone and the model
- Works in any app without per-app integrations, on a background screen that leaves the main screen usable
- Broad model choice, including local models on your network and ChatGPT, xAI or OpenRouter sign-in
- Payment steps are confirmed by default, and logins and verification codes are always handed to you
- Goes beyond app control with scheduled tasks, a sandbox for files and code, and MCP and CLI access from a computer
- API keys and memories are encrypted with the Android Keystore and conversations stay on the phone
Weaknesses
- Sideloaded APK plus Developer options, USB debugging and wireless-debugging pairing to operate apps
- Early beta from an individual developer, mainly tested on one phone
- Ask mode does not confirm sending messages, posting or deleting
- Model costs and quality depend on the provider you bring
- Android only, with the best experience on Android 14 or later
- Automating third-party apps can conflict with their terms of service
Mika pricing
Free (open source)
$0
- Full app under Apache-2.0, downloaded from GitHub releases
- No Mika account or server
- Bring your own API key, local model, or ChatGPT, xAI or OpenRouter sign-in
Note: Mika has no paid tier. Model usage is billed by the provider you connect, or counts against your ChatGPT Plus/Pro, xAI or OpenRouter plan or balance when you sign in with an account; local models on your own hardware have no per-call charge. The project notes that xAI's sign-in reuses the public entry of xAI's command-line tool and may be restricted, and that some xAI tiers cannot make calls after signing in.
Technical specs
Modalities
Text, Voice, Images
API pricing
Free; a local CLI task service and MCP server on the phone, reached over USB or a trusted Wi-Fi or private VPN. No hosted API.
Available models
Where Mika excels
Catching up on a busy group chat
Mika scrolls through the day's messages on its background screen and returns a summary, so you do not have to read hundreds of messages yourself or stop using the phone while it works.
Low-risk shopping research
It can filter products by price and rating and add the pick to the cart, while Ask mode stops before the order is placed and the payment itself stays with you.
Phone tasks from a coding agent
A computer-side agent can send a job to Mika over MCP or the CLI when the step only exists in a phone app, then wait for the result or answer Mika's questions.
Mika vs. competitors
Mika vs. Opposable
Opposable is a closed-source Android app (with an iPhone beta driven from a Mac) that exposes the phone to Claude Code, Codex and other MCP agents and includes a free on-device model, with paid plans for a hosted relay and cloud phones. Mika is Apache-2.0 open source, runs its own agent on the phone with a model you connect, works on a hidden background screen, and needs no vendor server, though it also accepts tasks from computer agents over MCP or a CLI.
Mika vs. Hark Pro
Hark Pro is a hosted personal agent for web, iOS and Android that runs errands on websites from its own cloud computer, on a free plan. Mika runs on your Android phone and operates your installed apps as well as websites, using your own model, with nothing hosted by a vendor.
Mika vs. ARTEMIS
ARTEMIS is Google's open-source Android test automation agent that runs on a computer and drives phones or emulators over ADB for QA and bug reproduction. Mika is an end-user assistant that runs on the phone itself and targets everyday tasks in your own apps.
Frequently asked questions
What is Mika?
Mika is an open-source AI agent app for Android. You tell it what you want by voice or text, and it opens the right apps or websites on a hidden background screen and taps, types and scrolls until the job is done, while you keep using your phone. It also handles scheduled tasks, questions about the screen you are looking at, and files in a built-in Linux sandbox.
Is Mika free and open source?
Yes. Mika is free and released under the Apache-2.0 license, with all code on GitHub. It does not include a model, so you pay your chosen model provider directly, use a local model, or sign in with a ChatGPT Plus/Pro, xAI or OpenRouter account that covers the usage.
Which phones does Mika support?
Mika installs on Android 11 or later, where it can chat, browse and work with files. Operating other apps needs Android 12 or later, and Android 14 or later is recommended because the background screen then has its own focus. The Linux sandbox needs an arm64 device. There is no iPhone version, and the app is installed as an APK from the project's GitHub releases.
Does Mika need root or special permissions?
Root is not required. To operate apps you turn on Developer options and USB debugging and pair Android's wireless debugging once, which lets Mika start a built-in helper and then switch on its own accessibility service. Phones without wireless debugging can start the helper with an adb command from a computer over USB, and rooted phones can start it with root instead.
Which AI models does Mika work with?
The main model must support tool calling. Mika has presets for providers including DeepSeek, Qwen, Kimi, Zhipu, MiniMax, OpenRouter, Anthropic, Gemini, OpenAI, xAI, Mistral, Groq, Ollama and LM Studio, and a Custom option for any OpenAI-compatible, Anthropic, Gemini or OpenAI Responses endpoint, including a model running on your own computer on the same network.
Does Mika ask before it pays or sends messages?
In the default Ask mode, Mika asks before placing orders or paying and shows where, what and how much; you complete the payment yourself. It does not ask before sending messages, posting, submitting forms, deleting or changing settings. Autonomous mode removes the payment confirmation. In both modes logins, verification codes and payment passwords are handed to you, and scheduled tasks in Ask mode skip money-spending steps.
Where does Mika store my data and API keys?
There is no Mika server. Your requests, screen contents and screenshots go only to the model provider you configure. API keys, memories and your profile are encrypted with the Android Keystore on the phone, conversations and run records stay on the device, and update checks only contact GitHub.
Can Claude Code or other computer agents use Mika?
Yes. AI tools on a computer can send tasks to Mika through a CLI with an agent Skill or through Mika's MCP server, over USB or over a trusted Wi-Fi or private VPN connection that is off by default. Payment confirmations still have to be handled on the phone.
Mika vs Opposable: what is the difference?
Both are Android apps that let an AI operate your signed-in apps. Opposable is closed source and centered on exposing the phone to Claude Code, Codex and other MCP agents, with a free on-device model, a paid hosted relay, an iPhone beta and rentable cloud phones. Mika is Apache-2.0 open source and works as a standalone phone assistant with your own model, a background screen, scheduled tasks and a sandbox, with no vendor server.
Integrations & fit
About Mika
Most AI assistants on phones answer questions or call a short list of app integrations. Mika instead works the phone the way a person does: it opens an app on a separate virtual display, reads the screen, and taps, types and scrolls until the job is done, while you keep using the main screen and can open a small live view to watch. Typical requests are summarizing a group chat, reading what a creator posted recently, adding a product to a cart and stopping before checkout, or setting several alarms. There is no per-app integration, though built-in text 'usage knowledge' for apps and sites such as WeChat, WhatsApp, Telegram, Gmail, Uber, Amazon and Google Maps makes common flows faster, and a built-in desktop-mode browser handles websites with persistent sign-ins. Around that core sit scheduled and conditional tasks ('check the weather at 8 and tell me only if it will rain'), a floating capsule that sends whatever screen you are looking at to Mika, hold-to-talk voice input, explicit memory, and a per-conversation Linux sandbox that runs Python, Node and shell on files you send. AI tools on a computer can also hand tasks to the phone through a CLI and Skill or an MCP server, and since early November 2026 Mika can itself connect to external MCP servers. There is no Mika server: you add your own API key for providers such as DeepSeek, Qwen, Kimi, OpenAI, Anthropic or Gemini, point it at Ollama or LM Studio on your network, or sign in with a ChatGPT Plus/Pro, xAI or OpenRouter account, so model costs are yours and screen contents go only to the provider you choose. By default it runs in Ask mode, which confirms only steps that spend money and hands logins, verification codes and payment passwords to you; sending messages, posting, deleting and changing settings are done without asking. The tradeoffs are significant: the APK is installed from GitHub releases, operating apps needs Android 12 or later (14+ recommended), Developer options with USB debugging left on and a one-time wireless-debugging pairing (or adb over USB, or root), and the project is an early public beta led by an individual developer that its own disclaimer says has mainly been tested on one Android 14 phone. Automating third-party apps may also breach their terms of service.
Updates from Mika
Mika v0.1.0 was published on GitHub as the first public beta, with app operation on a background screen, a built-in browser, screen sharing from a floating capsule, scheduled and conditional tasks, and a local Linux sandbox. Near-daily releases followed.
v0.1.32 added a page listing each running task's background screen with a live view, waiting when two tasks need the same app, a list of apps Mika must not touch (such as banking), and support for a self-deployed relay to reach the phone's CLI and MCP from a computer elsewhere (off by default).
v0.1.34 lets Mika connect to remote MCP servers or local ones in its sandbox, with account authorization and tool allowlists, shows token usage per turn and per conversation, and adds a configurable per-task step limit.
v0.1.37 lets rooted phones start Mika's phone helper with root, without wireless-debugging pairing or a computer. Root commands are a separate setting, off by default and behind a high-risk confirmation; in Ask mode, root commands that change or delete something need your approval.
v0.1.38 added manual creation and editing of one-off, interval, daily and weekly plans with reminders, conditions, cooldowns, time zones and idle-only runs, plus more reliable text input in hidden editors.
Are you the founder? Claim this listing →