AndroidHarness

Introduction: Android Harness AI Coding Agent Native app for mobile, No Root No PC
More: Author   ReportBugs   OfficialWebsite   
Tags:
AndroidHarness logo

An AI coding agent and chat assistant that runs on your Android phone.

No PC needed. No root needed.

Download from GitHub Get it on Google Play Buy me a coffee Join our Discord

Debug build Version Android 8.0+ License: MIT

About

AndroidHarness is a native Android app, written in Kotlin with Jetpack Compose, for working on real code projects from your phone. The agent reads and edits files, runs commands in a real Linux toolchain (bash, git, python, node), pushes to GitHub, drives a browser, and shows every step in a chat you can inspect.

The app has two sections:

  • Code is the agent. It works inside a project folder with files, shell, git, tests and browser tools.
  • Chat is for everyday questions, research, writing and brainstorming. It uses read-only tools and can hand work over to Code when a conversation turns into a project.

You can start for free with the built-in provider, which needs no account, or run models fully offline on the device.

Screenshots

Code agent building an app
Code agent at work
Chat answering a question with a rich card
Chat with rich answers
Provider picker
Switch providers
Approval card for installing Linux packages
Approve before installs
Code chat options menu
Chat tools and modes
Settings screen
Settings
Coding settings
Coding settings
Plugin manager
Plugins and skills
9Router local router page
9Router on the phone
Local models page
Offline local models

Highlights

  • A real coding agent. File edits, shell, git, tests, browser control and subagents. Every tool call gets a card you can open to see exactly what happened.
  • A real shell. Commands run through Shizuku, Wireless ADB pairing, optional root, a Termux-based Linux toolchain, or a toybox fallback. A shell policy and a secret redactor sit on top.
  • Free to start. The built-in keyless provider works without sign-up, and local GGUF models run offline. Cloud API keys and ChatGPT accounts are optional.
  • Free speech. Edge speech is built in with no key, more than 300 Microsoft voices and a sample for each. Groq, OpenAI, ElevenLabs, Deepgram, Cartesia and AssemblyAI can listen, speak or both.
  • Many providers. Anthropic, Google Gemini, ChatGPT accounts, any OpenAI-compatible endpoint, and fallback chains that switch models when one fails.
  • Research that reads. Turn on any of ten search providers; the agent picks one, falls through failures, reads the top pages and returns passages. Shaky statements can be flagged, and a question can be re-run deeper.
  • Plugins, skills and MCP. Install plugins from npm, GitHub, Claude marketplaces, archives or a folder on the phone, and the agent can manage them when you ask. Plugin commands, binaries and language servers run like Claude Code, and MCP servers connect over stdio or HTTP.
  • Built for long runs. A foreground service, resume cards, a saved message queue and optional auto-continue keep work safe when the screen turns off or the connection drops.
  • Safe by default. The agent stays inside the workspace, sensitive actions ask for approval, phone control is approved once per chat while destructive changes ask every time, and every change can be reviewed as a diff and undone.
  • 21 languages. Switch the app language at any time without a restart.

What's new in 4.0

  • Voice that is free and yours to choose. Edge speech needs no key and reads Microsoft's live catalogue of more than 300 voices. Groq, OpenAI, ElevenLabs, Deepgram, Cartesia and AssemblyAI can listen, speak or both, every voice has a sample button, and Auto send can be toggled from the + menu.
  • Phone control you approve. Agents ask once per chat before using Shizuku, root or Device Owner. Read-only checks run quietly, destructive changes ask every time, and Chat can set reminders, alarms and timers with approval.
  • Research that reads. Ten search providers can be switched on together. web_search reads the top pages and returns passages, shaky claims get a Check this claim marker, Dig deeper re-runs a question, and past chats can be placed where the answer refers to them.
  • Photos in chat. Paste images into the composer, see pending attachments in one tray, send every photo on the newest message, and open sent photos as a swipeable gallery.
  • A new look. Expressive settings, the AH monogram that acts out what the agent is doing, a tonal coding feed, a starter grid for new chats, and motion blur on every long list.
  • More agent control. The agent can manage plugins and reach more settings and skills, plugin commands, binaries, language servers and hooks work like Claude Code, and you can build your own settings sections.

See CHANGELOG.md for the full release notes.

Getting started

  1. Install the app from Google Play or GitHub Releases, or build it from source.
  2. Grant storage access. On Android 11 and newer the app needs All files access so the shell and file tools can use real paths.
  3. Pick a model. Use the built-in keyless provider, add an API key in Settings, or download a model in Settings → Local models.
  4. Recommended: install Shizuku, pair Wireless ADB in Settings → Terminal & device, or install the Linux environment from inside the app. This lets shell commands run with proper permissions. Rooted devices can also turn on direct root there.

Then open a workspace and ask for a change, or open Chat for anything that does not need a project. If something feels off, run /doctor to test every tool family.

Features

A short tour of what the app can do. The complete list, with every detail, is in FEATURES.md.

Code

  • Streaming chat with thinking blocks, tool cards, diffs and verification cards that list every build or test that ran.
  • A workspace is one folder, and a project groups workspaces for chats that need several, each editable or read only.
  • Plan mode, dual planning (plan on one model, execute on another), subagents and a persistent message queue that can hold send to choose when a queued message is read.
  • Approve phone control once per chat, set reminders, alarms and timers, and schedule a message from the chat itself.
  • Scheduled automations that run prompts daily, hourly or on demand.

Chat

  • Web research across ten providers, with the top pages read for passages, cited sources, claim checks and adjustable depth.
  • Reminders, alarms and timers, photo paste, a single attachment tray and a full-screen gallery.
  • Personal Chat agents with their own instructions, model, tools and knowledge files.
  • Rich answers with tables, charts, diagrams, images and interactive previews.
  • Shared memory with Code, saved answers and voice input and output.

Files, editor and Git

  • File manager, tabbed code editor with syntax highlighting, search and replace, and offline language tools.
  • Side-by-side diff viewer, per-chat change tracking with undo, and a Build & Test dashboard.
  • GitHub sign-in, repository import, commit and push, and pull requests from the app.
  • SSH workspaces and FTP or SFTP sync.

Models

  • Built-in free provider, ChatGPT accounts, Anthropic, Gemini and any OpenAI-compatible API.
  • Up to 20 rotating API keys per provider, proxies, custom headers, key testing and fallback chains.
  • Voice providers for listening and speaking, including free Edge speech with more than 300 voices.
  • On-device GGUF models with optional GPU acceleration, plus 9Router as a local model router.

Personalization

  • Custom color palettes, a background canvas, experimental glass effects and haptic feedback.
  • Expressive settings in named sections, custom settings sections you build yourself, and motion blur on long lists.
  • Phone and desktop layouts with keyboard shortcuts.

Shell access

Commands always run in the strongest environment the device allows.

Command target Runs as Environment
Privileged commands shell uid via Shizuku system access, like adb shell
Privileged commands, no Shizuku shell uid (2000) via Wireless ADB system access, like adb shell
Explicit root requests UID 0 via the verified root service full device access, when root is enabled
Workspace commands app uid Linux toolchain with bash, git, python and node
Fallback app uid toybox sh when nothing else is installed

Root is always optional and opt-in. More detail is in FEATURES.md.

FAQ

Do I need a PC? No. Editing, shell, git, GitHub, SSH, the models, and even building and signing APKs all happen on the phone.

Do I need an API key? No. The built-in provider offers free models without an account, and local models run offline. Keys are only needed for cloud providers.

Do I need root? No. Shizuku or the built-in Wireless ADB pairing gives shell-level access without unlocking the bootloader. Root is supported if you have it.

Does it work offline? Local models do, once downloaded. Web search, page reading and GitHub need a connection.

Which languages are supported? English, Simplified Chinese, Traditional Chinese, Russian, Arabic, Persian, Brazilian Portuguese, German, Spanish, Japanese, Indonesian, Sinhala, Vietnamese, Turkish, Malay, Thai, French, Hindi, Korean, Tamil, Bengali and Filipino.

Where do my chats and keys go? Chat traffic goes only to the provider you choose. Credentials are stored encrypted on the device, and secrets are redacted before they reach a model. Settings backups include API keys only if you ask, and ChatGPT sign-ins are never included.

Can it damage my files? The agent is limited to the workspace, every write shows up as a diff, and changes can be rewound or undone.

Which devices are supported? Android 8.0 and newer. Local models need a compatible 64-bit device.

Build from source

You need JDK 17 and the Android SDK.

./gradlew :app:assembleDebug
adb install -r app/build/outputs/apk/debug/app-debug.apk

Run the unit tests:

./gradlew :app:testDebugUnitTest

Debug APKs are also built on every push to main and nightly in the Actions tab. See CONTRIBUTING.md for the full setup.

GitHub setup

In Settings → GitHub, choose Sign in with GitHub or Use a personal access token.

  • Device sign-in shows a code that you enter on GitHub in your browser.
  • Only enable workflow access if you need to edit GitHub Actions files.
  • Fine-grained tokens need the target repositories and Contents: read and write. Editing workflows also needs Workflows access.
  • Organizations may require approval or SSO authorization.

Importing a repository needs the Linux environment. Clones go into private app storage, so broad storage permission is not required. Pushes never force-push, and each pushed commit is checked against the remote branch.

If a push automation fails, open its history to see the real cause. Common ones are the wrong workspace, missing Git, an SSH remote, missing token permissions, newer commits on the remote, or branch protection. A failed push can be retried with Push existing commits.

Building with your own OAuth app

Register a GitHub OAuth app, turn on Enable Device Flow, and set GITHUB_CLIENT_ID in the ignored local.properties file. Local debug and release builds include that value. The Client ID is not stored in the repository source, docs or public workflows, and no client secret or auth server is used. Builds without it still support token login. Keep a token as a backup, since GitHub rate limits Device Flow for shared OAuth apps.

Documentation

Contributing

Bug reports, fixes, tests and documentation are all welcome. Start with CONTRIBUTING.md and use the issue forms for bugs and feature requests. Please follow the Code of Conduct. Translations are handled on POEditor.

Acknowledgements

AndroidHarness takes ideas and design inspiration from these open source projects:

Hermes Agent, Aider, pi (ohmypi), OpenCode, Claude Code, Roo Code, Cline, browser-use, Termux, Shizuku, CodeGraph, Caveman, llama.cpp, sora-editor and Eruda.

License

Released under the MIT License.

Apps
About Me
GitHub: Trinea
Facebook: Dev Tools
AI Daily Digest