Mirroir MCP

<p align="center"> <img src="website/public/mirroir-wordmark.svg" alt="mirroir-mcp" width="128" /> </p>

mirroir-mcp

Let your agent use your real iPhone: errands across your apps, automations, even games.

<p align="center"> <a href="https://mirroir.dev/?utm_source=github&utm_medium=readme&utm_campaign=header">Website</a> · <a href="https://mirroir.dev/?utm_source=github&utm_medium=readme&utm_campaign=header#demo">Demos</a> · <a href="docs/onboarding.md">Set up with your agent</a> · <a href="docs/tools.md">Tools</a> · <a href="https://discord.gg/u8ug9QCSqF">Discord</a> </p> <p align="center"> <img src="website/public/recipe-grocery-list.gif" alt="The Claude app and a real iPhone side by side: the agent opens a saved Instagram recipe reel, listens to it, then writes an eleven-item grocery list in Reminders" width="720" /> </p> <p align="center"><em>A saved recipe reel becomes a grocery list in Reminders. The agent listens to the video and reads its frames. A real iPhone, sped up.</em></p>

No Developer Mode, no Xcode, no WebDriverAgent, no jailbreak. mirroir-mcp works through macOS iPhone Mirroring, so it reaches any app on your phone, including App Store apps you could never build. It works from any MCP client, with 42 tools.

It uses the apps already signed in on your phone: no connector, no OAuth, no password handed to anyone.

What people build

• A saved recipe reel becomes a grocery list. The agent opens the reel, reads its frames with describe_screen, hears the narration with listen, then types the ingredients into Reminders. That is the run in the GIF above.
• Errands across apps. Read the commute time in Waze, then post it in Slack: the agent carries what one app shows into another. Watch the demo. The same errand as a skill (Waze to Messages) is in Skills, exploration and CI.
• Games. A scene: step plays a turn-based or tap-driven game, frame by frame, under a hard frame, token and time budget; a run that runs out of budget is inconclusive (exit 66), never a silent pass. touch holds a finger for virtual joysticks, hold_keys drives games that take keyboard and mouse, and pinch and rotate zoom and rotate. iPhone Mirroring exposes a single pointer, so twin-stick games stay out of reach. See Games.

<p align="center"> <img src="website/public/chess-vs-maximum.gif" alt="The console and a real iPhone side by side during a full chess game: each move is logged as it is played, ending in checkmate against the Maximum bot on Chess.com" width="720" /> </p> <p align="center"><em>A real game on a real iPhone: the model reads the board, Stockfish picks the moves, mirroir plays them against Chess.com's strongest bot. The opening runs at 2.5×, the middle game at 45×, the finish at 2×.</em></p>

Quick start

You need macOS 15 or later and an iPhone connected through iPhone Mirroring. Paste this into your coding agent (Claude Code, Codex, Cursor, OpenClaw, …):

Set up mirroir-mcp so you can drive my iPhone: follow https://github.com/jfarcand/mirroir-mcp/blob/main/docs/onboarding.md step by step, and finish by reading back what my iPhone screen shows.

Or install it yourself:

npx -y mirroir-mcp install    # or: brew tap jfarcand/tap && brew install mirroir-mcp
npx -y mirroir-mcp doctor     # names anything left to fix

OpenClaw, Hermes Agent, Claude Desktop, Cursor, VS Code and the other clients: Install.

Then ask for something:

Open Calendar, create a new event called "Dentist" next Tuesday at 2pm.

Open my Expo Go app, tap "LoginDemo", test the login screen with
test@example.com / password123. Screenshot after each step.

Start recording, open Instagram, scroll the feed and open my profile, stop recording.

Build your own agent

• Any MCP client. Claude Code, Codex, Cursor, VS Code, Claude Desktop and the others connect to the same 42 tools. It also installs as a Claude Code plugin, an OpenClaw skill from ClawHub, or a Hermes Agent plugin that asks before every call that can change the phone: see Install.
• Proven steps in one call. Once a sequence has worked, run_steps replays it in one call ("tap: <text on screen>", "type: calculator", "wait_for: Get"). Every step is checked against the permission policy before the first runs, and the run stops at the first failure.
• Skills. Capture a workflow you repeat as a SKILL.md: numbered steps the AI follows, matched on screen text, not coordinates. generate_skill explores an app and writes them for you. Ready-made skills live in mirroir-skills; the format is in Skills Marketplace.
• Agents keep notes. save_note keeps one line per lesson per app, and get_notes gives it to the next session.
• An agent on another machine. mirroir-mcp serve speaks MCP over HTTP on 127.0.0.1 only, behind a bearer token; reach it through an SSH port forward. install-agent keeps it running as a LaunchAgent from its own signed "mirroir serve" bundle, so its grants survive upgrades.

Why mirroir

• Nothing to install on the phone. The iPhone stays stock: no signing, no provisioning profile, no on-device agent. If iPhone Mirroring shows it, mirroir can drive it.
• Your Mac stays yours. Taps, typing and swipes reach the mirror without bringing it to the front, so you can keep working. Drags and two-finger gestures bring it forward for a moment, then hand your app back.
• It hears the phone, too. listen turns what the phone plays, a reel's narration or a voice message, into text. It records only iPhone Mirroring's sound, and macOS transcribes it on the Mac: the audio never leaves it. Needs macOS 26; since 0.47.0.

Testing and CI

The run an agent makes once can become a test that runs without it.

• Explore once, replay without an LLM. generate_skill explores an app and writes SKILL.md files. A compiled skill replays from the CLI with no AI in the loop (mirroir test), so CI runs are deterministic and cost no tokens.
• One test across phone, web and backend. mirroir-run is our companion test runner, a separate binary. One YAML scenario can drive the iPhone, a browser, HTTP endpoints and subprocesses, then compare what the page and the phone showed. See Web and iOS in one scenario.
• Drift fails loudly. mirroir-run has a third verdict beside pass and fail: DRIFT, exit code 65, when every assertion held but the output moved past its baseline. Your CI lane decides whether drift blocks a merge.

How it compares

	mirroir	phone-harness	mobile-mcp	Appium
Physical iPhone, no Developer Mode or signing	Yes — stock phone via iPhone Mirroring	Yes — stock phone via iPhone Mirroring	USB-connected, trusted device (see its README)	No — WebDriverAgent via Xcode, Developer Mode
App Store apps you can't build	Yes	Yes	Yes	Yes, on a provisioned device
Your own iPhone from any MCP client	Yes	No — its MCP server is cloud-only; local phones use its CLI	Yes	Through the separate appium-mcp
Replay without an LLM	Yes — compiled skills, mirroir test	Python scripts you write; no recorder or compiled replay	No — the agent drives every run	Yes — test code
Distinct drift verdict	Yes — DRIFT, exit 65	No	No	No — pass / fail
One scenario across iPhone + web + HTTP	Yes — mirroir-run	No	No	Via your own test code
Android	No — an emulator window can be driven only as a generic macOS window	Yes	Yes	Yes
iOS Simulator	No	No	Yes	Yes
Hosted cloud phones	No	Yes — Phone Harness Cloud	Yes — Mobile Next Cloud	Via third-party device clouds

phone-harness is the closest tool here: it also drives a stock iPhone through iPhone Mirroring, and it adds Android and hosted phones, which mirroir doesn't have. mobile-mcp and Appium cover Android and simulators too. mirroir goes further after the agent's first run: that run becomes a replay that needs no LLM, and a test that tells you when the output moves.

Safety

Giving an AI access to your phone demands defense in depth, so mirroir-mcp is fail-closed.

• Read-only until you say otherwise. Without a config file, only the 14 read-only tools are exposed. Every tool that changes something stays hidden from the client until you allow it in permissions.json.
• Sensitive apps can be blocked. blockedApps keeps the AI out of apps like Wallet or Banking, even when other tools are allowed.
• Passwords stay yours. mirroir never presses a sheet that asks for a password or Touch ID, and it tells agents to ask you before sending, buying, deleting or changing a setting.
• No root, no telemetry, no daemon unless you install one. It runs as your user with Screen Recording and Accessibility, and lives and dies with your MCP client; only the opt-in serve agent keeps running, on loopback, behind a token. The one network call it makes by default is a once-a-day version check against npm (updateCheck: false turns it off). What your agent reads from the screen goes to whichever model your client uses.
• Kill switch. Quit iPhone Mirroring, unlock the iPhone, lock the Mac's screen, or quit your MCP client: each stops all input at once, and nothing reopens the app for the agent.

See Security for the full threat model.

Documentation

Install	Every install path, per-client setup, updating, uninstalling
Onboarding	Setup written for the agent: prerequisites, install, doctor, a read-only proof, and an optional first errand
Tools Reference	All 42 tools, parameters, and input workflows
How mirroir sees the screen	The observe, reason, act loop; OCR, icon detection, AI vision
Skills, exploration and CI	APP.md, skills, generate_skill, mirroir test, compiled replay, AI diagnosis, patterns
mirroir-run	Replay across iPhone, web, HTTP and subprocesses; GitHub Actions; Maestro import
Games	Shell and scene, the scene: step, game controls
CLI Tools	Recorder, doctor, keyboard layout, serve over HTTP
Configuration	All settings: screen intelligence, input timing, exploration, AI providers
FAQ	Security, focus stealing, keyboard layouts, embacle/vision mode
Security	Threat model, kill switch, and recommendations
Privacy	What mirroir reads, sends and stores
Permissions	Fail-closed permission model and config file
Known Limitations	Focus stealing, keyboard layout gaps, autocorrect
Patterns & Skills	Element patterns, screen recipes, APP.md app descriptions, and the detection pipeline
Exploring a New App	Step-by-step playbook for onboarding a new app — APP.md, permissions, components, exploration goal
YOLO Icon Detection	Recommended YOLO models, CoreML setup, and configuration
Compiled Skills	Zero-OCR skill replay
Testing	FakeMirroring, integration tests, and CI strategy
Web and iOS in one scenario	How mirroir-run hands iOS blocks to mirroir-mcp, and how cross_surface compares a browser with a phone
Cross-surface replay	mirroir-run scenario grammar, .mirroir/ plan, SAMPLE.md, judge profiles, CI
Troubleshooting	Debug mode and common issues
Contributing	How to add tools, commands, and tests
Skills Marketplace	Skill format, plugin discovery, and authoring

Community

Join the Discord server to ask questions, share skills, and discuss ideas. Contributions are welcome: see CONTRIBUTING.md. By submitting a patch, you agree to the Contributor License Agreement; your Git commit metadata serves as your electronic signature.

---

mirroir is developed by Jean-François Arcand at Async-IO. Apache 2.0 - Copyright 2026 Async-IO.org.

Why "mirroir"? — It's the old French spelling of miroir (mirror). A nod to the author's roots, not a typo.