Mobile app developers testing UI flows with an AI coding assistant
Phonedriver extends Claude Code, Codex, Cursor, or any MCP client to cover the mobile layer. Give the agent a task like 'go through the sign-up flow and report every bug you find with steps to reproduce' and it taps, reads, and reports from the actual device or emulator.
Skip if:
Your testing needs are covered by unit and integration tests, or your app is web-only. Phonedriver's value is on-device UI testing; if you are not doing that, you do not need it.
QA engineers automating device interaction through an MCP client
The 14 MCP tools cover the full tap, type, swipe, scroll, press, and launch surface a manual tester uses. Batch calls chain multiple steps in one round trip, and the CLI exposes the same commands for scripting outside of an AI context.
Skip if:
Your team already has a mature Appium or XCUITest suite. Phonedriver is better for exploratory or AI-driven testing than for running a large existing test corpus.
Developers demoing AI-driven phone tasks with an approval step
The approval guard stops the agent before payment actions. The demo on phonedriver.dev shows Claude Code ordering a pizza and stopping at Buy with GPay for the user's decision. This makes it safe to run live demos where payment actions appear: the agent cannot proceed without the user's explicit answer.
Skip if:
Your demo requires logging into an account protected by biometric auth. Phonedriver cannot pass Face ID, fingerprint checks, or a device passcode.
AI coding assistants can read and edit code, run terminal commands, and browse the web, but they stop at the mobile boundary. Testing a sign-up flow, reproducing a crash on a physical device, or walking through an app's onboarding in dark mode all require a developer to pick up a phone and do it manually. Automation tools like adb start a fresh process for every command and return raw XML accessibility trees, which is slow and token-heavy when an AI model is driving it. Cloud testing platforms can provide real devices, but they add per-session costs, require an account, and send device interactions through a third-party server.
The result is that AI-assisted development effectively ends at the web layer. Teams either skip mobile UI testing in their AI workflows, or they write brittle scripts that break when a screen changes.
Instead of sending raw screenshots to the model, Phonedriver reads the device's accessibility tree and returns a short numbered list of UI elements. After a tap, the reply shows only what changed, using + for appeared, - for left, and ~ for changed. On an Android emulator, reading the screen takes 13 ms against 2,164 ms with adb's uiautomator dump.
Payment and action approval guard
When the agent tries to tap a button labeled Pay, Buy now, Buy with GPay, Place order, Start subscription, Transfer, or Delete account, Phonedriver holds the tap and shows a prompt in the AI client. The model never sees the prompt and cannot answer it. The check reads labels in 11 languages: English, Spanish, Portuguese, French, German, Italian, Dutch, Japanese, Chinese (Simplified), Korean, and Hindi.
Android and iOS device support
Works with Android emulators and physical Android devices over ADB, and with the iOS Simulator and real iPhones and iPads over Xcode's XCTest. On Android there is no app to install on the phone. On iOS, `phonedriver ios setup` reads the signing team from your developer certificate, then builds, installs, and starts the XCTest runner. After setup, Wi-Fi connections are supported so the cable can come off.
14 MCP tools and a matching CLI
The MCP server exposes 14 tools: devices, apps, snapshot, screenshot, tap, type, swipe, scroll, press, launch, wait, batch, clipboard, and shell. The snapshot, screenshot, and wait tools are marked read-only so a client can run them without a permission prompt. The same commands are available through the phonedriver CLI for scripts or for agents that prefer a shell.
Batch operations and path tapping
A single batch call chains multiple steps: tap, type, press, wait, scroll, and swipe in one round trip. A single tap call can follow a path: `tap {"target":"More options","then":["Settings","Dark theme"]}` navigates three levels in one call. Models tap one step at a time by default; giving the agent the printed skill tells it to send paths.
Hidden secondary display on Android 13 and later
On Android 13 and later, Phonedriver can run the agent's work on a hidden second screen, leaving your main display untouched. Setting PHONEDRIVER_DISPLAY=hidden in the MCP config or passing --display hidden to the CLI activates this. Tested on an emulator and a HONOR X8 running Android 13.
6.9x fewer tokens than mobile-mcp on nine real phone jobsOn nine jobs run on the same Android emulator on 10 November 2026 with Claude Sonnet 5.5, Phonedriver used 658 thousand tokens and passed 27 of 27 runs. mobile-mcp used 4.6 million tokens and passed 17 of 18. phone-use used 1.1 million tokens and passed 14 of 18. The benchmark methodology, raw data, and dismissed runs are published in docs/bench/jobs.md.
No telemetry, connects only to your devicePhonedriver sends no telemetry. After the initial binary download from GitHub, the MCP server connects only to the attached device. A test suite runs the whole tool in a sandboxed environment that blocks every other network connection, so the no-telemetry claim is verified on every CI run.
AGPL-3.0, fully self-hosted, no per-session costPhonedriver is AGPL-3.0 licensed. You can run it on your own machine, use it commercially, and modify it. If you change it and offer it to others as a network service, you must publish your changes. There is no cloud account to create and no per-session cost.
170x faster screen reads than adbOn an Android emulator, Phonedriver reads the screen in 13 ms median versus 2,164 ms with adb's uiautomator dump, 170x faster. Screenshots take 9 ms versus 291 ms, 32x faster. The speed comes from a small helper process Phonedriver starts on the phone that keeps the accessibility connection open, where adb starts a new process and writes XML each time.
Trade-offs
-Real iPhone requires a Mac with Xcode and an Apple Developer certificateiOS Simulator works from any Mac with Xcode. A real iPhone or iPad additionally requires an Apple Development certificate. `phonedriver ios setup` automates the signing and installation, but the Mac and the developer account are prerequisites. Phonedriver does not support iOS on Windows or Linux.
-Cannot pass biometric auth or enter a passcodePhonedriver cannot unlock a device with Face ID, Touch ID, or a PIN. On a locked phone, a command stops and reports the device is locked. Runs that require an unlocked device must start from a device that is already unlocked, or use an emulator with no lock screen.
-New project launched November 2026, limited third-party coverageThe repository was created on 8 November 2026 and had 148 stars and 1 open issue at the time of this writing. The benchmarks and device coverage are the author's own. Third-party reviews, long-term community feedback, and coverage of less common Android devices are not yet available.
BrowserStack is a cloud platform for web and mobile testing with access to hosted real devices. Phonedriver connects to devices you own or run as emulators and operates as an MCP server for AI agents, so the overlap is narrow: both give access to a device for testing, but through different models.
Feature
Phonedriver
BrowserStack
License
AGPL-3.0
Proprietary
Device hosting
Local: your emulators and USB devices
Cloud: hosted real devices
AI agent interface
Native MCP server
No MCP integration
Pricing
Free (self-hosted)
Per-session or subscription
Phonedriver is the stronger choice when your team uses AI coding assistants and wants direct, low-latency access to a local emulator or USB device. The text-based screen reader produces 6.9x fewer tokens per job than mobile-mcp, which matters when token usage is a real cost. BrowserStack is the better option when you need a large matrix of physical devices you do not own, or when your existing test automation runs on Appium or Selenium and you want a hosted device lab without infrastructure management.
Phonedriver vs RobotActions
RobotActions is a hosted MCP service that gives AI agents control of real iPhones and Android devices in the cloud. Both tools expose an MCP interface for AI coding assistants; the choice is primarily about where device control runs.
Feature
Phonedriver
RobotActions
License
AGPL-3.0
Proprietary
Device hosting
Local (your machine)
Cloud (RobotActions servers)
Pricing
Free
Paid service
Data privacy
No data leaves your machine
Device interactions on third-party servers
Phonedriver is the better option for teams that cannot route device interactions through external servers, want no recurring cost, or need to test on devices they already own. RobotActions is an option when you want managed real devices without USB cables or a Mac for iOS testing.
Does Phonedriver work with iPhones, or only Android?
Phonedriver works with both. Android support covers emulators and physical devices over ADB, with no app required on the phone. iPhone and iPad support requires a Mac with Xcode and, for real devices (not the Simulator), an Apple Developer certificate. phonedriver ios setup handles the signing and installation automatically. After setup, the cable can come off: phonedriver ios wifi on switches the connection to Wi-Fi.
Can an AI agent using Phonedriver make payments or delete my account without asking me?
No, unless you explicitly allow it. Phonedriver holds any tap on a button labeled Pay, Buy, Buy with GPay, Place order, Start subscription, Transfer, or Delete account and shows a prompt in your AI client. The model cannot see or answer that prompt. You choose Allow, Allow for this session, or Decline. If you want the agent to run without asking, you can pass autonomous: true or set PHONEDRIVER_GUARD=off in the MCP config.
How does Phonedriver compare to mobile-mcp on token usage?
On nine Android jobs run on 10 November 2026 with Claude Sonnet 5.5, Phonedriver used 658 thousand tokens total and passed all 27 runs. mobile-mcp used 4.6 million tokens and passed 17 of 18 runs. phone-use used 1.1 million tokens and passed 14 of 18 runs in 22.7 minutes versus Phonedriver's 5.1 minutes. The benchmark data and job descriptions are at docs/bench/jobs.md in the repository.
What MCP clients does Phonedriver support?
Phonedriver is a standard MCP server over stdio, so any MCP client can run it. Tested clients include Claude Code, Codex, Cursor, VS Code, GitHub Copilot, Gemini CLI, Windsurf, Zed, Cline, and opencode. Each client uses either the claude mcp add command or a JSON entry in its MCP config file. Setup guides for each client are on phonedriver.dev.
Can I use Phonedriver at work or in a commercial project?
Yes. Phonedriver is AGPL-3.0 licensed. You can use it, run it internally, and use it in commercial products. If you modify it and offer a network service based on the modified version to others, you must publish your changes under the same license. The Android helper includes files from scrcpy under Apache-2.0, noted in NOTICE.