Ghost tapping a phone

Ghost in the Droid

Summon a ghost into your phone.
It sees the screen. It taps the buttons. It never sleeps.

Ghost driving nine real phones

_{▶ Watch full HD with sound on YouTube · Download mp4}

Nine agents. Nine real devices. One ghost. (Click to watch.)

The pitch, in one breath

Every AI agent can think. Almost none can touch a phone.

Ghost is the body. It gives any LLM agent a real Android phone or iPhone: it reads the screen, taps, swipes, and types through a clean tool surface, and it scales from one device on your desk to a whole phone farm. Point Claude Code, Codex, Antigravity, Cursor, your LangChain app, or a model running inside the phone itself at it, and your agent grows hands.

Bring your own brain. Keep the same body. MIT, forever.

Mix and match: the Ghost matrix

Ghost 1.3 is the only Android + iOS agent framework where platform, brain, and driver are all swappable. Pick one from each column. They all compose.

Pick a platform	Pick a brain	Pick a driver
Android over USB ADB	Claude Code (free with your Max/Pro sub)	MCP client (Claude Code, Codex, Antigravity, Cursor, Opencode)
Android over Wi-Fi ADB	Anthropic API	LangChain toolkit
Android in a Docker + KVM emulator pool	OpenRouter (any model, one key)	LlamaIndex tool spec
iPhone over Appium + WebDriverAgent	Ollama (fully local)	Ghost CLI (`ghost "book a table" --device pixel`)
iPhone over Tailscale + WebDriverAgent (wireless)	vLLM (your own GPU)	Web dashboard chat with a live phone stream
	On-device — llama.cpp (`.gguf`, any quant), MediaPipe (`.task`, Gemma mobile bundles), MLX (Apple Silicon, opt-in). The model runs in the phone, airplane-mode works	REST API (`/docs` OpenAPI)

_{The Docker + KVM emulator pool is Android-only; iOS uses Apple's own simulator and Xcode signing, which are macOS-native.}

The body stays the same 62 MCP tools no matter what you plug in. Every new model release is a free upgrade to your phone agent.

See it in action

Nine demos, one per feature, each recorded on real devices. Click any clip for the full mp4.

MCP clients: point your agent at a phone
_{Claude Code: type in the TUI, watch it drive the phone.}	_{Codex: OpenAI's CLI operating a real phone over MCP.}	_{Antigravity: Google's agent, same 62-tool body.}
Interfaces: CLI, framework, browser
_{Ghost CLI: ghost "check reddit" --device asus.}	_{LangChain: Ghost's tools in a framework agent.}	_{Web chat: chat in your browser, it drives a real phone.}
On-device and iOS: the phone as body and brain
_{On-device Android: model runs in the app, airplane mode.}	_{On-device iPhone: your iPhone talks to itself.}	_{iPhone, real app: Ghost does the Duolingo lesson.}

What It Does

Open-source Python framework for controlling Android and iOS devices from one agent harness. Android runs through ADB. iOS runs through Appium XCUITest and WebDriverAgent, with real iPhones as the target path and simulators for development and CI.

Define skills for any app, run them from the dashboard or API, scale across a phone farm.

The ghost taps what you'd tap

Android control through ADB: tap, swipe, type, clipboard, shell, intents, and stealth variants
iOS control through WebDriverAgent: screenshot, accessibility tree, tap, swipe, type, app launch, clipboard, browser actions
Live phone screen streaming: Android MJPEG/WebRTC, iOS WDA MJPEG with screenshot fallback
Interactive touch-to-tap on the streamed screen
Multi-device phone farm with per-device job queues

Forge reusable skills for any app

YAML-based UI element definitions per app
Platform-specific selectors with elements.yaml for Android and elements_ios.yaml for iOS
Python action classes with precondition checks
Multi-step workflows that chain actions together
Built-in skills for TikTok and Play Store
iOS browser/news demo skill and smoke-level TikTok iOS workflows
Skill Hub: browse, search, and install skills from the community registry
Install from CLI: android-agent skill install tiktok

Teach the ghost new tricks

BFS-based auto app explorer: discovers every screen and transition
LLM-assisted Skill Creator: chat with AI while viewing the live device stream
The AI identifies UI elements and generates action/workflow code

Scale the haunting

Multi-device phone farm with per-device job queues
Bot runner: queue, schedule, and monitor automation jobs
Per-device integration tests with Android screen recording or iOS WDA MJPEG recording

Platform Support

Surface	Android	iOS
Device ref	ADB serial, e.g. `emulator-5554`	`ios:<udid>`
Backend	ADB + optional Portal companion app	Appium XCUITest + WebDriverAgent
Real device support	Yes	Yes, with Mac/Xcode/WDA signing
Simulator/emulator support	Android emulator tooling	Booted iOS simulators via Appium/WDA
Live stream	Portal WebRTC, MJPEG, screencap	WDA MJPEG, screenshot polling fallback
Screen tree	Android UIAutomator XML	Normalized XCTest accessibility tree
Skills	`elements.yaml`, Android packages	`elements_ios.yaml`, iOS bundle IDs
Android-only today	ADB shell, intents, wireless ADB, Play Store helpers, Portal overlay	Unsupported with stable platform errors

Why it is built this way

Skills cost $0. Thinking costs tokens. An agent should not pay an LLM to tap a button it has tapped a thousand times. Known workflows compile to skills: deterministic YAML + Python recipes that replay in seconds with zero LLM calls. Use AI for the unknown task, replay for the known one.

The phone is the easiest sandbox there is. Give the ghost an old Android with its own SIM and its own accounts, physically separate from your life. You already know how to set up a phone.

Zero-app by default. A fresh install touches nothing on the Android device: screen reads go through uiautomator, actions through adb shell input, no root, no accessibility service. Want it faster? The optional Portal companion app gives roughly 30x quicker UI reads. Want it private? On-device mode runs the whole loop inside the phone, nothing leaves it. All opt-in.

Local-first, cloud-optional. The server and dashboard run on your machine. With a local or on-device brain, your screenshots never leave your network. Pick a cloud brain and prompts go to that provider, same as any tool. Your call, always.

Requirements

Python 3.10+
Android: Android phone with USB debugging enabled and ADB on PATH (adb devices should list your phone)
iOS: macOS with Xcode, Appium 2 + XCUITest driver, and a trusted iPhone or booted simulator
Node.js 18+ (for the frontend dev server)

iOS Support

Ghost drives iPhones through Appium/WebDriverAgent with the same tool surface: ios:<udid> device refs route tap/swipe/type/screenshot to WDA, and iOS-aware browser primitives (open_url, read_news, extract_visible_text) cover web tasks. Android-only tools (shell, launch_intent, Portal overlay) return a clear platform error instead of failing silently.

iOS is opt-in: enable it with GITD_ENABLE_IOS=1 (or ios_platform_enabled=true in .env). Wireless drive rides your Tailscale tailnet, so treat the tailnet as the trust boundary. See docs/SETUP_IOS.md for Appium/WDA setup, and on-device on iPhone to run the model on the iPhone itself.

Quick Start

Zero-install with uvx (or pipx):

bash

uvx ghost-in-the-droid doctor   # green/red preflight: python, adb on PATH, devices, ports, LLM keys
uvx ghost-in-the-droid login    # sign in with your Claude subscription, no API key needed
uvx ghost-in-the-droid up       # server + dashboard at http://localhost:5055  (API docs at /docs)

doctor prints a checklist with fix hints instead of a stack trace when something is missing (e.g. adb not on PATH). Prefer pipx? pipx install ghost-in-the-droid gives you the ghost-in-the-droid / android-agent commands.

Sign in with your Claude subscription (no API key)

If you have a Claude Max/Pro subscription, you don't need an API key. android-agent login signs you in through the claude CLI's own Anthropic OAuth flow and points Ghost at the claude-code provider:

bash

android-agent login       # opens Anthropic sign-in via the claude CLI

Ghost never handles or stores your token; the claude CLI owns it, refresh included. doctor shows a green Claude subscription check once you are signed in. To use an API key instead, set ANTHROPIC_API_KEY (or OPENAI_API_KEY / OPENROUTER_API_KEY) and pick that provider.

Run fully local with Ollama (no keys, no cloud)

Select Ollama in the Phone Agent tab. Runs entirely on your machine with Ollama:

bash

brew install ollama       # or: curl -fsSL https://ollama.com/install.sh | sh
ollama serve &
ollama pull llama3.2:3b   # 2GB, fast, good tool-use

From a clone (for development)

bash

git clone https://github.com/ghost-in-the-droid/android-agent.git
cd android-agent
pip install -e ".[all]"

android-agent doctor        # preflight
android-agent up            # start server + dashboard on :5055

# Frontend (separate terminal)
cd frontend && npm install && npx vite --host 0.0.0.0 --port 6175
# Dashboard at http://localhost:6175

iOS Quick Start

iOS support requires a Mac because Appium uses Xcode's XCUITest/WebDriverAgent stack. Real iPhones also require trust, Developer Mode, UI Automation permission when prompted, and WDA signing with an Apple development team.

server.ts

# 1. Install and run Appium XCUITest
npm install -g appium
appium driver install xcuitest
appium --base-path /

# 2. Find your iPhone or booted simulator UDID
xcrun xctrace list devices
xcrun simctl list devices booted

# 3. Configure the backend for iOS
export IOS_DEVICE_UDID="<udid>"
export IOS_APPIUM_URL="http://127.0.0.1:4723"
export IOS_BUNDLE_ID="com.google.chrome.ios"       # or com.apple.mobilesafari
export IOS_MJPEG_SERVER_PORT="9100"                # use unique ports per iOS device

# 4. Run a product-path smoke workflow
uv run python scripts/ios_chrome_news_smoke.py \
  --device "ios:<udid>" \
  --bundle-id "$IOS_BUNDLE_ID" \
  --url https://text.npr.org/ \
  --max-headlines 5 \
  --max-articles 3 \
  --fix-health \
  --out-dir data/ios_chrome_news_smoke

For full real-device signing, simulator, WDA MJPEG, health recovery, scheduler, and MCP setup details, see docs/SETUP_IOS.md.

Environment Variables

Copy .env.example to .env (if provided) or create a .env file in the project root. The server reads configuration via Pydantic Settings. Optional variables include:

Variable	Purpose
`OPENAI_API_KEY`	LLM features (Skill Creator, Agent Chat)
`ANTHROPIC_API_KEY`	Alternative LLM provider
`OPENROUTER_API_KEY`	OpenRouter LLM provider
`DEFAULT_DEVICE`	ADB serial (auto-detected if empty)
`IOS_DEVICE_UDID`	iPhone or simulator UDID; devices are addressed as `ios:<udid>`
`IOS_APPIUM_URL`	Appium server URL, default `http://127.0.0.1:4723`
`IOS_BUNDLE_ID`	Default iOS app/browser bundle, e.g. `com.google.chrome.ios` or `com.apple.mobilesafari`
`IOS_DEVICES_JSON`	Per-device iOS config for multiple phones/simulators, WDA ports, bundle IDs, and MJPEG ports
`IOS_MJPEG_SERVER_PORT`	WDA MJPEG stream port; use one unique port per iOS device

Give any AI agent a mobile body (MCP)

Ghost ships an MCP server with 62 tools for real-device control. Any MCP-compatible client can use them. Android serials receive the Android implementation; ios:<udid> refs route to the iOS backend where supported and return stable unsupported-platform errors for Android-only tools. One command wires it up:

bash

# Claude Code (same shape for Codex, Cursor, VS Code Copilot, Windsurf)
claude mcp add android-agent -- uvx --from ghost-in-the-droid android-agent-mcp

uvx installs the package, builds an isolated env, and runs the server. No clone, no venv.

Other clients (Codex, Claude Desktop, Cursor, VS Code, Windsurf)

bash

# Codex (OpenAI)
codex mcp add android-agent -- uvx --from ghost-in-the-droid android-agent-mcp

Claude Desktop (claude_desktop_config.json), Cursor (.cursor/mcp.json), Windsurf (mcp_config.json):

config.json

{
  "mcpServers": {
    "android-agent": {
      "command": "uvx",
      "args": ["--from", "ghost-in-the-droid", "android-agent-mcp"]
    }
  }
}

VS Code Copilot (.vscode/mcp.json) uses the same block under a "servers" key. Contributors who clone the repo get a ready .mcp.json, all 62 tools live on first claude launch.

The 62 tools, by category:

Category	What the ghost can do
See	`screenshot`, `get_elements`, `get_screen_tree`, `get_screen_xml`, `screenshot_annotated`, `screenshot_cropped`
Touch	`tap`, `tap_element`, `swipe`, `long_press`, `type_text`, `type_unicode`, `press_back`, `press_home`, `press_key`
Apps	`launch_app`, `search_apps`, `list_apps`, `launch_intent`, `force_stop`, `list_packages`
Understand	`get_phone_state`, `classify_screen`, `find_on_screen`, `ocr_screen`, `ocr_region`, `extract_visible_text`
Browser / iOS	`open_url`, `browser_back`, `get_current_url`, `read_news`, `extract_articles`, `wait_for_text`
Device	`list_devices`, `clipboard_get`, `clipboard_set`, `get_notifications`, `open_notifications`, `toggle_overlay`, `device_health`
Skills	`list_skills`, `run_workflow`, `run_action`, `create_skill`, `explore_app`
Batch	`run_flow` / `chain`: one call runs a whole multi-step recipe server-side, N round-trips collapse to 1
Diagnostics	`list_crashes`, `get_crash`, `web_search`, screen recording, camera + TTS

toggle_overlay, launch_intent, Android shell helpers, wireless ADB, and Play Store helpers remain Android-only and return stable platform errors on iOS.

Run the brain inside the phone

New in 1.3: the model can live on the device. The Ghost companion app embeds real inference engines, so the phone is both the body and the brain. Airplane mode on, agent still working, nothing leaving the device.

Platform	Engine	Model format	Best for
Android	llama.cpp (JNI)	`.gguf`	Any GGUF: Gemma, Llama, Mistral, Qwen, DeepSeek
Android	MediaPipe	`.task`	Small Gemma models, fastest Android integration
iPhone	llama.cpp (Metal)	`.gguf`	Same GGUF as Android, Qwen2.5 1.5B on Metal
iPhone	MLX	Apple Silicon	Faster decode on newer chips, opt-in

On iPhone, Qwen2.5 1.5B runs through llama.cpp on Metal and drives the phone's own UI, with an opt-in MLX engine for faster decode on Apple Silicon. Small models are kept honest with grammar-constrained decoding so tool calls always parse. Full detail on the on-device LLM page.

Skill Hub CLI

bash

android-agent skill search tiktok        # search the public registry
android-agent skill install tiktok       # install a skill
android-agent skill install github.com/someone/their-skill
android-agent skill list                 # what is installed
android-agent skill update tiktok        # update a skill
android-agent skill validate ./my-skill/ # check before publishing

The skill registry lives in registry/ in this repo. Community skills are auto-discovered nightly from repos tagged android-agent-skill.

Teach the Ghost a New App

Two ways to forge a skill:

Community skill (your own repo): create a repo with skill.yaml, elements.yaml, actions, and workflows, tag it android-agent-skill, and it appears on the Skill Hub automatically (nightly scraper).

Official skill (PR to this repo): build and test as a community skill first, then open a PR adding it to registry/. CI validates, a maintainer reviews, and it gets the "Official" badge.

Each skill needs:

skill.yaml: metadata (name, version, app package or iOS bundle ID, supported platforms, actions, workflows)
elements.yaml: Android UI element resource IDs and descriptions
elements_ios.yaml: optional iOS selectors for XCTest accessibility trees
actions/: Python classes extending Action with precondition() and execute()
workflows/: Python classes extending Workflow with steps()

Prefer to let AI do it? The Skill Creator watches a live device stream and generates the action and workflow code while you narrate. Full guide in CONTRIBUTING.md.

The dashboard

A local-first control room at http://localhost:5055 (or :6175 in dev). WebRTC and MJPEG live streams, tap-on-screen, tracing, and every tab below.

Tab	What it does
Phone Agent	Live device stream (MJPEG/WebRTC), tap/swipe on screen, multi-device view
Scheduler	Cron-style job scheduling with queue management
Skill Hub	Browse installed skills, run actions and workflows, export/delete
Skill Creator	LLM-assisted skill builder with live device stream
Skill Miner	Auto app explorer: BFS state discovery with screenshots
Manual Run	Start/stop bot jobs, queue management, logs
Tests	Per-device test runner with screen recording playback
Emulators	Create, boot, snapshot, and manage Android emulators
Tracing	Per-turn traces, token accounting, tool-call visibility

Architecture

Code

android-agent/
  run.py                    # Uvicorn entry point on :5055
  gitd/
    app.py                  # FastAPI app factory + plugin hook
    models/  schemas/       # SQLAlchemy 2.0 ORM + Pydantic v2
    routers/  services/     # Route handlers + business logic
    skills/                 # Skill packages (tiktok, play_store, safari [iOS], tiktok_ios)
    bots/common/adb.py      # Android Device class: tap, swipe, dump tree, wait_for
    bots/common/ios.py      # iOS Device class: Appium/WDA session, UI tree, gestures
    mcp_server.py           # MCP server: 62 tools for any LLM agent
  frontend/                 # Vue 3 + Vite + TypeScript + Tailwind
  portal/                   # Kotlin companion app (WebRTC, on-device inference)
  site/                     # Docs site (Astro + Starlight)

The flow: FastAPI backend on :5055 exposes device control, skills, bots, scheduling, and streaming. The Vue SPA talks to it over /api/*. Skills define per-app interaction (elements + actions + workflows). Device backends route by ref: bare serials use bots/common/adb.py, ios:<udid> refs use bots/common/ios.py and Appium/WDA. State lives in SQLite via SQLAlchemy 2.0 and Alembic.

Benchmark (sneak peek)

Early result, with the full methodology writeup still cooking: driven by Claude Code, Ghost completes 115 of 116 tasks (99.1%) on AndroidWorld, Google Research's benchmark for Android agents, on the unmodified upstream harness. Treat this as a preview rather than a citable number. The detailed writeup and trajectories are coming. Watch the releases.

Tech Stack

Layer	Technology
Backend	FastAPI, Uvicorn, Python 3.10+
Data	SQLAlchemy 2.0, Alembic, Pydantic v2, SQLite (WAL)
Frontend	Vue 3, TypeScript, Vite, Tailwind CSS 4
Device control	ADB for Android; Appium XCUITest/WebDriverAgent for iOS
Streaming	Android MJPEG/WebRTC via Portal; iOS WDA MJPEG plus screenshot fallback
On-device	llama.cpp, MediaPipe (Android); llama.cpp on Metal, MLX (iOS)
Quality	Ruff, pytest, Playwright

Running Tests

Most tests are unit/API tests and run without a live device. Live Android and iOS integration tests require the relevant device stack.

bash

# Run all tests on a specific device
DEVICE=<serial> python3 -m pytest tests/ -v --tb=short

For iOS live smoke tests:

bash

IOS_LIVE_NEWS_TEST=1 \
IOS_DEVICE_UDID="<udid>" \
IOS_APPIUM_URL="http://127.0.0.1:4723" \
IOS_BUNDLE_ID="com.google.chrome.ios" \
uv run --extra test python -m pytest tests/test_browser_news.py::test_live_ios_chrome_news_workflow

Get your Android device serial from adb devices.

Database Migrations

The project uses Alembic for schema migrations:

bash

alembic revision --autogenerate -m "add new_field to my_table"   # after editing a model
alembic upgrade head                                              # apply pending
alembic downgrade -1                                              # rollback one

Contributing

The ghost gets stronger with every skill. See CONTRIBUTING.md for adding app skills (highest impact), writing actions and workflows, backend architecture, and the PR process. Join the community on the Skill Hub.

License

MIT. The ghost is free. The ghost is open source. The ghost is yours.