Agent

An autonomous, always-on agent that is yours alone. It has its own browser, keeps a wiki memory of everything it learns about you, and hands the heavy jobs to your own machines.

It lives in two places, and you can use either or both: on your phone, or on your computer with the phone as the window. Same agent, same memory, same approvals; the heavy work and the keys stay where you put them.

You bring the model. Paste an API key, or sign in with the ChatGPT or Claude plan you already pay for.

Get the APK | Get the desktop app | Build it from source

Free and open source, MIT. No account, no subscription, no server of ours.

Home, chat, memory, routines. The two demo recordings are on the project page.

It is not on the Play Store. You sideload the APK.

1. Open Releases and download the latest Agent-*.apk on your phone.
2. Tap the file and allow the install when Android asks. That is the whole sideload.
3. Open Agent, give it a name, pick how it looks, then pick how to power it. See Setup.
4. On a Samsung, Xiaomi or OnePlus phone, set the app's battery to Unrestricted under Settings, or Android will kill the browser and the routines while they work.

You need Android 8 or newer on an arm64 phone, which is every phone sold in the last few years. The APK is about 36 MB without the bundled subscription runtimes, and about 360 MB with them, so leave room for it and for its data.

On the phone. The agent loop, the browser, memory, tasks and routines all run in the app. Nothing else is needed, and no other computer has to exist.

On your computer, with the phone as the window. Pair the phone with the desktop app (Settings, Devices, Add a device). From then on the phone shows the same chats, cards, approvals and reports, and the work runs on the computer with its accounts, its files and its bigger keys. It behaves like one agent, not two:

• The phone knows which computer is answering. A message goes to the brain that should take it; if that computer says nothing within a few seconds, the next one gets the same message id, so the work still runs exactly once.
• A card is only "working" while the computer that published it is answering. A stopped machine says so instead of spinning.
• Your computer's files stay on your computer. When the agent shows a video or a picture, the phone asks for exactly the part of the file it needs, by a token the computer made for that one card, and nothing else is reachable.

It goes and looks. The agent drives its own built-in browser. It cannot control your existing Chrome tabs, so you sign in to your sites inside it. That browser runs in a foreground service, so it keeps working while you are in another app or the screen is off. When a site throws a captcha or a sign-in at it, it hands the page to you and carries on when you are done. Profiles keep Personal and Work apart, and helper agents can split a big job up.

It remembers you. Everything it learns gets a page in a wiki, linked to the pages around it. Read it, change it, pin it or delete it. Nothing is kept that you cannot see.

It answers in cards, not walls of text. A table, a chart, a set of steps, a row of buttons, one card per listing: the answer is something you can read at a glance, tap, and edit. Pictures and videos the agent finds or makes play inside the chat.

It runs on a clock. A routine can fire at a set time, on a new email, or the moment an app you chose sends a notification. It can watch a price for weeks and only speak up when it actually moves. A routine works with the same tools the agent has, including your task list.

It learns the web apps you use. Show it a site once and it can write its own tools for that site, so the next visit is one call instead of a hundred clicks. You decide whether those tools may act without asking.

It builds little apps for you. A mini-app is a small dashboard or tracker it makes and keeps up to date, on the phone and on the computer, with only the permissions you grant it.

It uses the accounts you have. Apps connect through Composio with your own key: Gmail, Calendar, Drive, Slack and a few hundred more. It asks to connect one only when a job needs it, and the sign-in opens inside the app. You can hold several accounts per app, like two Outlooks or four Gmails.

It hands heavy work to a real computer. Any server or PC you can SSH into becomes a machine it can use. Run a command, start a long job that keeps going in the background, move files both ways. Looking around just runs. Anything that changes the machine shows you the exact command first.

It answers in voice, and reads your files. Voice notes in and out, live calls if you add a Gemini key, photos and documents in chat, including scanned PDFs read on the device.

It tidies up overnight. At 02:00, and only while charging and online, it sorts its memory, writes notes from the day and builds your morning screen.

It waits for your yes. Anything that spends money or sends something in your name stops at an approve button.

• Not a product. No account, no subscription, nobody to pay.
• Not on the Play Store, so nobody at Google has looked at it. Read the source, or build it yourself, if that matters to you.
• Not finished. It has mostly been tested on one Samsung phone, and the phone-and-computer pairing is the newest part.
• Not faultless. It is an agent with a browser and access to whatever you connect, so it will get things wrong. Read what you approve.
• Not for everyone yet. It expects you to be comfortable sideloading an APK and pasting a key.

Nothing is read until you allow it. Location, files, notifications and the rest are asked for at the moment a job needs them, and the app runs with all of them denied. The one most people want is notifications, and even then you pick the apps it may read, one at a time, and take it back whenever you like.

• Keys are stored on the phone in EncryptedSharedPreferences, locked by the Android Keystore.
• Memory, tasks, routines and chat live in a Room database on the phone.
• Traffic leaves the phone only to the model you picked, the sites the agent browses, Composio if you added a key, your voice provider if you added one, and the computers you paired.
• Machine secrets stay encrypted on the phone and the agent only ever sees a machine's name. A machine is pinned the first time you connect, and refused if its identity changes later.
• A file the agent shows you is served by the computer that holds it, to the paired phone, by an opaque token with no path in it, read-only. A made up token gets nothing.

If an agent that lives on your own hardware and answers to you is something you want to keep existing, give the repo a star. It is the only signal this project has, and it is what puts Agent in front of the next person looking for it.

Onboarding walks you through this. In order:

1. Name and style. Name your agent, pick a colour and a face. The face becomes the app's icon.

2. Power. Choose one:

   Option	What you need	Notes
   Your own computer	The Agent desktop app with Devices turned on	The phone becomes the window; nothing to pay, keys stay on the computer
   API key	A key from DeepSeek, OpenRouter, Gemini, Claude, OpenAI or OpenCode (Zen or Go)	Runs in the app's own agent loop. Quickest to set up.
   Any OpenAI-compatible endpoint	Base URL, model name and a key	Hosted services, or your own server. See Self-hosted servers.
   ChatGPT subscription	A ChatGPT account (the free tier worked in testing)	Runs Codex on the phone through the bundled runtime. Sign in inside the app.
   Claude subscription	Claude Pro or Max	Runs Claude Code on the phone through the bundled runtime. Read Risks first.

   Once a key is checked you can pick the model, and switch from the chip at the top of the chat at any time.

3. Connections, if you want them. Paste your Composio API key under Connections, then connect the apps you use. The agent can also ask for one mid-job, and that sign-in opens inside the app's browser.

4. Voice, if you want it. Under Settings, Voice, add a Gemini, OpenAI or ElevenLabs key for spoken replies. A Gemini key also turns on live calls.

5. Notifications, if you want them. Choose which apps the agent may read. Nothing is read until you pick.

Agent speaks English and Spanish (written by people) and 53 more languages, translated by machine. Pick one in Settings > You > Language, or in Android's own per-app language page (Android 13 and later). "Phone's language" follows the phone. The machine translated ones say so in the picker.

English, Spanish (Magarto), and, machine translated: Afrikaans (Afrikaans), Amharic (አማርኛ), Arabic (العربية), Bengali (বাংলা), Catalan (Català), Chinese, Simplified (中文 (简体)), Chinese, Traditional (中文 (繁體)), Czech (Čeština), Danish (Dansk), Dutch (Nederlands), Filipino (Filipino), Finnish (Suomi), French (Français), German (Deutsch), Greek (Ελληνικά), Gujarati (ગુજરાતી), Hausa (Hausa), Hebrew (עברית), Hindi (हिन्दी), Hungarian (Magyar), Igbo (Igbo), Indonesian (Bahasa Indonesia), Italian (Italiano), Japanese (日本語), Kannada (ಕನ್ನಡ), Khmer (ខ្មែរ), Korean (한국어), Malay (Bahasa Melayu), Malayalam (മലയാളം), Marathi (मराठी), Nepali (नेपाली), Norwegian (Norsk bokmål), Persian (فارسی), Polish (Polski), Portuguese (Brazil) (Português (Brasil)), Portuguese (Portugal) (Português (Portugal)), Punjabi (ਪੰਜਾਬੀ), Romanian (Română), Russian (Русский), Sesotho (Sesotho), Swahili (Kiswahili), Swedish (Svenska), Tamil (தமிழ்), Telugu (తెలుగు), Thai (ไทย), Tswana (Setswana), Turkish (Türkçe), Ukrainian (Українська), Urdu (اردو), Vietnamese (Tiếng Việt), Xhosa (isiXhosa), Yoruba (Yorùbá), Zulu (isiZulu).

The model's own instructions and tool descriptions stay in English on purpose; only what you read on screen is translated.

Machine translation gets things wrong. If you speak one of these languages, please fix it with a pull request, the way @magarto did Spanish, whose structure all the languages follow:

1. Edit the files in app/src/main/res/values-<language>/. The file names and the keys are the same as in values/ (English).
2. Keep the placeholders (%1$s, %d) and the HTML tags as they are, and escape apostrophes as '.
3. Run python3 scripts/check-translations.py. It checks keys, placeholders, HTML, escaping, plural forms for the language, and leftover English. The unit tests run the same check.
4. When a language has been proofread by a native speaker, send a PR that sets machine = false for it in AppLocale.kt, and the "machine translated" line goes away.

Right to left languages (Arabic, Hebrew, Persian, Urdu) mirror the whole app. To add a language that is not here, copy values/ to values-<code>/, translate it, and add one line to AppLocale.choices.

Pick Other (OpenAI-compatible), then enter the server's base URL, usually ending in /v1, the model name and a key. Any placeholder works if your server does not check keys.

• HTTPS is the default and the way to do it. Put the server behind a reverse proxy with a certificate such as Caddy or nginx, or use tailscale serve to get an https:// address on your tailnet.
• Plain HTTP on your LAN (http://192.168.1.20:8080/v1 and the like) is blocked unless you opt in. When the URL starts with http:// you get a warning and a checkbox, Allow unencrypted HTTP for this server. Requests and your API key then travel in clear text, so only tick it on a network you trust. The opt-in is saved with that server and cleared when you switch provider.
• Built-in providers always use HTTPS.

Your phone can be the remote, not the engine. Under Connections, Machines you can add any computer you can SSH into: a VPS, a university or work server, your own PC.

• Sign in with a password, a private key you already use, or a new key made on the phone. You paste its public half into ~/.ssh/authorized_keys once, and no password is stored.
• Through a gateway, so a machine behind a login host can be reached as well.
• Commands that change things ask first. The card shows the exact command, unless you switch on Run without asking for that machine.
• At home, put the phone and the computer on the same Tailscale network and use the computer's tailnet address. On Windows, turn on OpenSSH Server under Settings, Optional features, or run sshd inside WSL.

The agent talks to the other apps on your phone the way Android apps talk to each other, with intents, and other apps can talk to it the same way. Everything stays on the phone.

Two tools do this.

• android_apps shows what is installed and what each app accepts: links, places, shared text and images, and its exported screens. It only reads, so it never asks.
• android_intent sends an intent: {kind: broadcast | activity | service, action, package, component, data, mimeType, category, extras, flags}. Extras keep their type (text, true or false, numbers, lists), or you force one with {type, value}.

A Google Maps route, for example:

{kind: "activity", action: "android.intent.action.VIEW", data: "google.navigation:q=Sandton+City", package: "com.google.android.apps.maps"}

What asks first:

• An intent that only opens a screen goes straight away: view a link or a place, open an app, the dialler with a number filled in, a search, a settings page.
• Anything else asks you first: sharing into an app, a broadcast, a service.
• To stop being asked for one, add its action and app to Settings, Automation, Agent talks to your apps, Allowed app intents, one per line: android.intent.action.SEND | com.whatsapp. Use * for any app. The approval card shows the exact line.
• When nobody is there to answer, as in a routine at night, it is not sent.

Android does not let an app open a screen from the background. If the Agent app is not on screen, the intent arrives as a notification you tap. Allow Display over other apps for Agent to skip that.

macrodroid_run is a preset over the same thing: one broadcast your macro listens for.

1. In MacroDroid, make a macro with the trigger Intent Received.
2. Action: com.past9.phoneaos.MACRO, exactly. It is case-sensitive.
3. Add one extra. Name: macro, in small letters. Value: the macro's name, for example demo, with nothing after it, not even a space.
4. Optional: v1, v2, v3 and params = *, each saved to a variable. The agent always sends all of them as text.
5. In the Agent app, open Settings, Automation, Agent talks to your apps, and add demo to Allowed macros. Macros on that list run without asking. Any other macro asks first.
6. Tap Send under Test a macro. It sends the first macro on your list.

If the macro does not run, MacroDroid gives no error. Check these, in order:

• MacroDroid is switched on, and so is the macro.
• The extra's name is macro, not Macro. Phone keyboards capitalise it. The agent also sends Macro and MACRO, so this one is covered from 0.15.32.
• The value has no space after it. Predictive keyboards add one, and then it never matches. demo* also works.
• Remove the extra and test with the action alone. If that fires, the extra was the problem.

Tasker catches the same broadcast: Profile, Event, Intent Received, Action com.past9.phoneaos.MACRO. The values arrive as %macro, %v1, %v2, %v3 and %params.

The MacroDroid webhook is optional and off by default. It lets an agent that is not on this phone reach MacroDroid's own server, trigger.macrodroid.com. Your values then pass through MacroDroid's server, so leave it off unless you need it.

Two doors, with the same token and the same switch (Settings, Automation, Automation from other apps, Allow outside runs). A run started from outside has nobody watching, so anything that needs your approval is refused.

• By intent. Send a broadcast to package com.past9.phoneaos with action com.past9.phoneaos.RUN_ROUTINE and the extras token and routine (its name or id), plus an optional note. Or action com.past9.phoneaos.INSTRUCT with token and text. This works when the Agent app is closed. The package is required, because Android does not deliver a broadcast to a closed app without it. The sheet has a recipe to copy, with your token in it.
• Over HTTP. http://127.0.0.1:47321 with the token as a Bearer header, only while the app is running. This one can also read the result back. Steps: docs/automation-macrodroid.md.

The agent on a linked computer can use android_intent, android_apps and macrodroid_run on the phone through the device link, with the same allowlists and the same question on the phone. Switch it off under Settings, Linked devices.

WhatsApp Web calls, in-browser editors and models that run in the page need shared memory (SharedArrayBuffer), which Android's WebView hides unless the app names the site. Settings, Browser, Sites that need advanced browser features is your list: https://web.whatsapp.com to start with, exact https sites only, no wildcards. In the same sheet, Let the agent speak in WhatsApp calls (off by default) lets the agent's live voice go into a WhatsApp Web call it places: place_whatsapp_call places a whole call and returns what was said, browser_call_voice does the steps by hand. Turn on calling for a test account (off by default, with a ban-risk warning) is for a test number WhatsApp has not given calling to, never your main number. Details and the check to run on the phone: docs/browser-advanced-features.md.

You need JDK 17, the Android SDK (compileSdk 37), and for the runtime pack curl, ar, patchelf and Node/npm.

git clone https://github.com/Past-da-king/agent
cd agent
# Downloads Termux's Node build, Codex and ripgrep and packs them as jniLibs plus assets,
# which are git-ignored. Only needed for the subscription paths. API keys work without it.
./tools/build-runtime-pack.sh
./gradlew assembleDebug
adb install -r app/build/outputs/apk/debug/app-debug.apk

Release builds are signed from a properties file that is never in this repository: pass -PsigningProps=<file>.

Tests: ./gradlew testDebugUnitTest, which runs the unit tests plus Robolectric and Roborazzi screenshots in light and dark.

• Kotlin and Jetpack Compose (Material 3 Expressive), Room, WorkManager.
• On the API key path, a provider-neutral agent loop in Kotlin that speaks both chat formats.
• On the subscription path, Node 24 (Termux's Android build) ships inside the APK as lib*.so files, because Android only lets an app execute files from its native library folder, and the Agent SDK runs on it. Codex and Claude Code ship the same way as their arm64 musl builds; Claude Code gets the musl loader as libldmusl.so, and a local proxy resolves DNS for both. The app's own tools are served to them over a local MCP server on 127.0.0.1 with a per-install token.
• The browser is an Android WebView hosted in a foreground service.
• With a computer paired, the phone and the computer replicate chats, tasks and memory over an encrypted link of their own: no relay of ours, no account.

• Your own risk. The agent can browse, fill in forms, read the notifications you allow and act in the apps you connect. Approvals guard payments and sends, so check what you approve.
• The Claude subscription login is not officially sanctioned. The option is there because it works. Using it is a call you make about your own account. A Claude API key is the safe way to use those models here.
• Sideloaded. It is not reviewed by Google.
• Battery. Background browsing and routines use battery. The overnight routine only runs while charging.
• Early software. Tested mainly on one Samsung phone.

This repository started as an early version of the app and carries the work of its contributors. The line that has been shipping since was developed privately and was merged into main in November 2026, keeping the earlier history and the licence intact. See Releases for what changed when.

MIT. See LICENSE.