mirroir-mcp

Give your AI eyes, hands, and a real iPhone. An MCP server that lets any AI agent see the screen, tap what it needs, and figure the rest out — through macOS iPhone Mirroring. Experimental support for macOS windows. 33 tools, any MCP client.

Requirements

• macOS 15+
• iPhone connected via iPhone Mirroring

Install

/bin/bash -c "$(curl -fsSL https://mirroir.dev/get-mirroir.sh)"

or via npx:

npx -y mirroir-mcp install

or via Homebrew:

brew tap jfarcand/tap && brew install mirroir-mcp

The first time you take a screenshot, macOS will prompt for Screen Recording and Accessibility permissions. Grant both.

Claude Code

claude mcp add --transport stdio mirroir -- npx -y mirroir-mcp

GitHub Copilot (VS Code)

Install from the MCP server gallery: search @mcp mirroir in the Extensions view, or add to .vscode/mcp.json:

{
  "servers": {
    "mirroir": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "mirroir-mcp"]
    }
  }
}

Cursor

Add to .cursor/mcp.json in your project root:

{
  "mcpServers": {
    "mirroir": {
      "command": "npx",
      "args": ["-y", "mirroir-mcp"]
    }
  }
}

OpenAI Codex

codex mcp add mirroir -- npx -y mirroir-mcp

Or add to ~/.codex/config.toml:

[mcp_servers.mirroir]
command = "npx"
args = ["-y", "mirroir-mcp"]

git clone https://github.com/jfarcand/mirroir-mcp.git
cd mirroir-mcp
./mirroir.sh

Use the full path to the binary in your .mcp.json: <repo>/.build/release/mirroir-mcp.

How it works

Every interaction follows the same loop: observe, reason, act. describe_screen gives the AI every text element with tap coordinates (eyes). The LLM decides what to do next (brain). tap, type_text, swipe execute the action (hands) — then it loops back to observe. No scripts, no coordinates, just intent.

Beyond the basics, the server exposes higher-level navigation and lifecycle tools: press_back navigates back by OCR-tapping the "<" chevron (with a canonical-position fallback), press_home/press_app_switcher/spotlight drive system gestures, scroll_to scrolls until a target text becomes visible, reset_app force-quits an app via the App Switcher, set_network toggles connectivity through Settings, and measure times how long an action takes to surface a target element. Multi-target setups use list_targets/switch_target to move between window automation endpoints. Skill authoring is served by record_step and save_compiled (compiled replay) and calibrate_component (test a component definition against the live screen). See the Tools Reference for all 33 tools.

Describe Your App

mirroir can explore any iOS app blindly, but it works better when you tell it what to expect. Write an APP.md file and mirroir reads it before exploration starts:

---
app: Santé
archetype: dashboard
obstacle_mode: auto
---

## Structure
Dashboard with 4 tabs: Résumé, Partage, Parcourir, Profil.

## Résumé Tab
- Summary cards for health metrics that drill down to charts
- Cards often show "Aucune donnée" on test devices

## Obstacles
- Health Access permission → tap "Autoriser"
- Notification permission → tap "Ne pas autoriser"

## Skip
- Supprimer les données de Santé
- Réinitialiser

What the code actually uses today: archetype overrides recipe auto-detection; obstacles are auto-dismissed when obstacle_mode: auto; skip merges with permissions.json.skipElements; tabs (inline or as a section) are injected as high-priority targets; Structure + tab body + Tips become AI context in generated skills.

See the APP.md specification for the complete field list, loader resolution rules, and the permission-system bridge. Three levels of patterns work together — elements (what rows look like), screens (what the page layout means), and apps (what the developer knows). Patterns & Skills covers the full system.

Examples

Paste any of these into Claude Code, Claude Desktop, ChatGPT, Cursor, or any MCP client:

Open Messages, find my conversation with Alice, and send "running 10 min late".

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 Settings, scroll to General > About, stop recording.

Screen Intelligence

describe_screen is the AI's eyes. Three backends work together to give the agent a complete picture of what's on screen — text, icons, and semantic UI structure.

Apple Vision OCR (default)

The default backend uses Apple's Vision framework to detect every text element on screen and return exact tap coordinates. This is fast, local, and requires no API keys or external services.

Icon Detection (YOLO CoreML)

Text-only OCR misses non-text UI elements — buttons, toggles, tab bar icons, activity rings. Drop a YOLO CoreML model (.mlmodelc) in ~/.mirroir-mcp/models/ and the server auto-detects it at startup, merging icon detection results with OCR text. The AI gets tap targets for elements that text-only OCR cannot see.

AI Vision Mode (embacle)

Instead of local OCR, describe_screen can send the screenshot to an AI vision model that identifies UI elements semantically — cards, tabs, buttons, icons, navigation structure — not just raw text. This produces richer context for the agent, especially on screens with complex layouts.

The embacle runtime is embedded directly into the mirroir-mcp binary via Rust FFI. describe_screen calls the embedded runtime in-process — no separate server, no network round-trip, no additional setup. The FFI layer (EmbacleFFI.swift → libembacle.a) handles initialization, chat completion requests, and memory management across the Swift/Rust boundary.

embacle routes vision requests through already-authenticated CLI tools (GitHub Copilot, Claude Code) so there is no separate API key to manage. If you have a Copilot or Claude Code subscription, you already have access.

Install

brew tap dravr-ai/tap
brew install embacle          # CLI tools (embacle-server, embacle-mcp)
brew install embacle-ffi      # Rust FFI static library (libembacle.a)

Then rebuild mirroir-mcp from source (or reinstall via Homebrew) so the binary links against libembacle.a:

# From source
swift build -c release

# Or via Homebrew (rebuilds automatically)
brew reinstall mirroir-mcp

Zero-config activation

When the embacle FFI is linked into the binary, screenDescriberMode defaults to "auto" which automatically resolves to vision mode. No settings change required — install embacle-ffi, rebuild, and describe_screen starts using AI vision.

To force local OCR even when embacle is available, explicitly set "ocr":

// .mirroir-mcp/settings.json
{
  "screenDescriberMode": "ocr"
}

See Configuration for all available settings.

Skills

When you find yourself repeating the same agent workflow, capture it as a skill. Skills are SKILL.md files — numbered steps the AI follows, adapting to layout changes and unexpected dialogs. Steps like Tap "Email" use OCR — no hardcoded coordinates.

Place files in ~/.mirroir-mcp/skills/ (global) or <cwd>/.mirroir-mcp/skills/ (project-local).

APP.md

Describe your app's structure to guide exploration — see Describe Your App above and the full APP.md specification. Place APP.md files in ~/.mirroir-mcp/skills/ or the mirroir-skills repo at patterns/apps/.

---
version: 1
name: Commute ETA Notification
app: Waze, Messages
tags: ["workflow", "cross-app"]
---

## Steps

1. Launch **Waze**
2. Wait for "Où va-t-on ?" to appear
3. Tap "Où va-t-on ?"
4. Wait for "${DESTINATION:-Travail}" to appear
5. Tap "${DESTINATION:-Travail}"
6. Wait for "Y aller" to appear
7. Tap "Y aller"
8. Wait for "min" to appear
9. Remember: Read the commute time and ETA.
10. Press Home
11. Launch **Messages**
12. Tap "New Message"
13. Type "${RECIPIENT}" and select the contact
14. Type "On my way! ETA {eta}"
15. Press **Return**
16. Screenshot: "message_sent"

${VAR} placeholders resolve from environment variables. ${VAR:-default} for fallbacks.

Skill Marketplace

Install ready-to-use skills from jfarcand/mirroir-skills:

git clone https://github.com/jfarcand/mirroir-skills ~/.mirroir-mcp/skills

From Exploration to CI

The generate_skill tool lets an AI agent explore an app and produce SKILL.md files. It uses breadth-first search (BFS) to traverse the app as a navigation graph — screens are nodes, tappable elements are edges. The explorer describes each screen, matches elements against component definitions to decide what to tap, visits child screens, and backtracks via the back chevron. Duplicate screens are skipped via structural fingerprinting. See Component Detection below for how the explorer interprets raw elements into structured UI components.

The explorer works viewport-by-viewport: after calibrating the page length, it builds a plan from the current viewport, taps elements top-to-bottom, scrolls down to reveal more content, and rebuilds the plan for each new viewport. This approach works with both OCR and AI vision describers. Pass seed for deterministic ordering across runs.

Exploration is bounded — it does not discover every reachable screen in large apps. Depth, screen count, and time limits keep runs practical. For targeted flows, provide a goal to focus the traversal.

graph TD
    A["Launch App"] --> B["Describe Screen"]
    B --> C{"Calibrated?"}
    C -- No --> D["Scroll Full Page"]
    D --> E{"skip_calibration?"}
    E -- No --> F["Component Detect +\nClassify + Validate"]
    E -- Yes --> G["Classify Elements\nDirectly"]
    F --> H["Build Plan"]
    G --> H
    C -- Yes --> H

    H --> I{"Untried\nElements?"}
    I -- Yes --> J["Tap Element"]
    I -- No --> K["Return to Root"]

    J --> M["Describe +\nClassify Edge"]
    M --> N{"Transition"}
    N -- new screen --> O["Add to Frontier"]
    O --> P["Backtrack"]
    N -- revisited/dead --> P

    P -- push: tap back --> H
    P -- modal: tap close --> H
    P -- tab: tap prev --> H

    K --> Q{"Frontier\nEmpty?"}
    Q -- No --> R["Next Frontier\nScreen"]
    R --> B
    Q -- Yes --> S["Generate SKILL.md"]

Generate

Two modes: autonomous exploration (BFS) and guided session (manual step-by-step).

Autonomous BFS exploration — the agent explores on its own:

Explore the Settings app and generate a skill that checks the iOS version.

This calls generate_skill(action: "explore", app_name: "Settings", goal: "check iOS version") under the hood. The explorer launches the app, runs BFS from the root screen, and outputs a SKILL.md for the discovered path.

Guided session — the AI navigates manually, capturing each screen:

1. generate_skill(action: "start", app_name: "MyApp") — launch app, OCR first screen
2. Use tap/swipe/type_text to navigate, then generate_skill(action: "capture") to record each screen
3. generate_skill(action: "finish") — assemble captured screens into a SKILL.md

Test

Run skills deterministically from the CLI — no AI in the loop:

mirroir test apps/settings/check-about
mirroir test --junit results.xml --verbose        # JUnit output
mirroir test --dry-run apps/settings/check-about    # validate without executing

Exit code 0 = all pass, 1 = any failure.

By default the CLI auto-recompiles a compiled skill whose screen fingerprint has drifted; --no-auto-recompile disables that and reuses the stale coordinates.

Compiled Skills

Compile a skill once to capture coordinates and timing. Replay with zero OCR — a 10-step skill drops from 5+ seconds of OCR to under a second.

mirroir compile apps/settings/check-about        # compile
mirroir test apps/settings/check-about            # auto-detects .compiled.json
mirroir test --no-compiled check-about            # force full OCR

AI agents auto-compile skills as a side-effect of the first MCP run. See Compiled Skills for details.

AI-Assisted Diagnosis

When a test step fails, pass --agent to get an AI diagnosis of what went wrong and suggested fixes:

mirroir test --agent gpt-5.3 apps/settings/check-about
mirroir test --agent claude-sonnet-4-6 apps/settings/check-about
mirroir test --agent ollama:llama3 apps/settings/check-about
mirroir test --agent embacle apps/settings/check-about

Built-in agents:

Custom agents are defined as YAML profiles in <cwd>/.mirroir-mcp/agents/<name>.yaml (project-local) or ~/.mirroir-mcp/agents/<name>.yaml (global, project-local wins). Pass the file's <name> to --agent.

A profile runs in one of two modes:

mode: api — call a cloud or local HTTP endpoint:

# ~/.mirroir-mcp/agents/my-gpt.yaml
mode: api
provider: openai          # anthropic | openai | ollama | embacle
model: gpt-5.3
api_key_env: OPENAI_API_KEY
base_url: https://api.openai.com
system_prompt: "You are a terse iOS automation debugger."
max_tokens: 4096

mode: command — run a local CLI process (e.g. an already-authenticated claude -p / copilot -p, or a custom script). The diagnostic payload is delivered two ways: if any args entry contains ${PAYLOAD}, the JSON payload is substituted there; otherwise it is piped to the command's stdin:

# ~/.mirroir-mcp/agents/claude-cli.yaml
mode: command
command: claude
args: ["-p", "${PAYLOAD}"]
system_prompt: "Diagnose the failed step and reply with JSON."

embacle routes requests through already-authenticated CLI tools (GitHub Copilot, Claude Code, etc.) — no separate API key needed:

brew tap dravr-ai/tap && brew install embacle
mirroir test --agent embacle my-skill

Pattern System

The explorer uses a three-layer pattern system to understand iOS apps — the same declarative concept at different scales:

• Element patterns (patterns/elements/) — 34 definitions matching row-level UI components (table rows, tab bars, toggles, summary cards). Each specifies match rules, interaction behavior, and grouping logic.
• Screen patterns (patterns/screens/) — 7 archetype recipes that identify screen-level navigation models from element composition. Auto-detected during calibration, or declared via archetype in APP.md.
• App patterns (patterns/apps/) — APP.md files with structure, obstacles, skip lists, and archetype declarations. The developer's source of truth.

Built-in archetypes: settings-list, dashboard, social-feed, content-grid, conversation-list, utility-display, detail-form.

Place custom patterns in ~/.mirroir-mcp/components/ (elements), ~/.mirroir-mcp/recipes/ (screens), or the mirroir-skills repo.

Vision Indicators

AI vision describers describe UI elements semantically ("Activité chevron") rather than character-by-character ("Activité" + ">"). A vision-indicators.md file maps these descriptions to OCR-compatible characters so the component pipeline works identically with both backends:

## Indicators
- chevron: >
- dismiss: ×
- back: <

When a vision element ends with a mapped suffix (e.g. "Entraînements chevron"), the normalizer splits it into two elements: "Entraînements" + ">". Place vision-indicators.md alongside your component definitions.

See Component Detection for the full definition format, match rule reference, and the detection pipeline.

Replay anywhere with mirroir-run

mirroir-mcp captures iOS flows — AX + OCR + BFS exploration. mirroir-run replays .mirroir/ SkillStep scenarios on Linux CI against web, process, and HTTP surfaces. Both speak one SkillStep grammar, so an iOS capture and a web scenario are the same language on two surfaces — tied together by cross_surface equivalence rather than maintained as two bespoke suites. A single Rust binary (runner/ in this repo), independent of the Swift server.

Drop a .mirroir/ directory in any repo and mirroir-run discovers it from the working directory:

your-app/
└── .mirroir/
    ├── mirroir.yaml          # the plan: must-pass / nice-to-pass entries
    └── apps/<sample>/
        ├── SAMPLE.md         # how to boot the app under test
        ├── APP.md            # structure, obstacles, skip lists
        └── scenarios/*.yaml  # SkillStep flows to replay

cd your-app
mirroir-run                          # discover .mirroir/, replay the must-pass plan
mirroir-run --scenarios all          # include nice-to-pass entries too

Each plan entry either points at a local: sample tree or extends a shared archetype (archetypes: ["<pack>/<name>@<ver>"]) with per-instance vars: and boot:. An archetype captures a reusable app shape — say, an AI chat console — once, and parameterizes it per app.

Where the iOS leg comes from. generate_skill … emit=true (on finish or explore) writes the captured flow into .mirroir/apps/<app>/ as a --validate-able iOS scenario plus a cross-surface oracle (baselines/<flow>.ios.txt) and a cross_surface parity gate — additive to the web leg's runnable scenarios. Run the MCP from your consumer repo (or pass output_dir) so the tree lands in the right .mirroir/; emitting into ~/.mirroir (the runner's pack home) is refused. Pair it with a web capture and mirroir-run asserts the two surfaces stay equivalent. See runner/docs/mirroir-dotfile.md for the pairing convention.

Web steps compile to a Playwright .spec.ts and run across chromium, firefox, and webkit. Selectors resolve three ways: raw CSS, Playwright engine prefixes (role=button[name="Save"], text=Welcome, xpath=…), or data-test + visible text. Process and HTTP steps dispatch natively; an LLM judge step scores agent responses against expected signals, and drift detection catches output divergence vs. a baseline.

Install mirroir-run

# crates.io
cargo install mirroir-run

# Homebrew
brew install jfarcand/tap/mirroir-run

Prebuilt binaries for macOS (Intel + Apple Silicon), Linux (gnu + musl), and Windows are attached to each runner-v* release. See runner/docs/ for the scenario grammar, SAMPLE.md schema, judge profiles, and CI integration.

Security

Giving an AI access to your phone demands defense in depth. mirroir-mcp is fail-closed at every layer.

• Tool permissions — Without a config file, only the 11 read-only tools (screenshot, describe_screen, start_recording, stop_recording, get_orientation, status, check_health, list_targets, list_skills, get_skill, calibrate_component) are exposed. Every mutating tool — including press_back — is hidden from the MCP client entirely, so it never sees them.
• App blocking — blockedApps in permissions.json prevents the AI from interacting with sensitive apps like Wallet or Banking, even if mutating tools are allowed.
• No root required — Runs as a regular user process using the macOS CGEvent API. No daemons, no kernel extensions, no root privileges — just Accessibility permissions.
• Kill switch — Close iPhone Mirroring to kill all input instantly.

// ~/.mirroir-mcp/permissions.json
{
  "allow": ["tap", "swipe", "type_text", "press_key", "launch_app"],
  "deny": [],
  "blockedApps": ["Wallet", "Banking"]
}

See Permissions and Security for the full threat model.

CLI Tools

Recorder

Record interactions as a skill file:

mirroir record -o login-flow.yaml -n "Login Flow" --app "MyApp"

Doctor

Verify your setup:

mirroir doctor
mirroir doctor --json    # machine-readable output

Configure

Set up your keyboard layout for non-US keyboards:

mirroir configure

Updating

# curl installer
/bin/bash -c "$(curl -fsSL https://mirroir.dev/get-mirroir.sh)"

# npx
npx -y mirroir-mcp install

# Homebrew
brew upgrade mirroir-mcp

# From source
git pull && swift build -c release

Uninstall

# Homebrew
brew uninstall mirroir-mcp

# From source
./uninstall-mirroir.sh

Configuration

All settings live in settings.json — project-local (.mirroir-mcp/settings.json) or global (~/.mirroir-mcp/settings.json). Project-local settings override global ones. Every setting also has a corresponding environment variable (e.g. MIRROIR_SCREEN_DESCRIBER_MODE).

{
  "screenDescriberMode": "auto",
  "agent": "embacle",
  "ocrBackend": "auto",
  "keystrokeDelayUs": 15000,
  "explorationMaxScreens": 30
}

See Configuration Reference for all 40+ settings covering screen intelligence, input timing, scroll behavior, exploration budgets, AI providers, and keyboard layouts.

Documentation

Community

Join the Discord server to ask questions, share skills, and discuss ideas.

Contributing

Contributions welcome. By submitting a patch, you agree to the Contributor License Agreement — your Git commit metadata serves as your electronic signature.

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