ios-agent-driver

An MCP server that lets an AI agent drive the iOS Simulator in a loop — so an agent can actually use your app: tap, type, swipe, read the screen, and verify what happened.

It bridges the gap between iOS development and agentic testing. The primitives to control a simulator exist (xcrun simctl, Meta's idb), but nothing packages them into tools an agent can call to close the perceive → decide → act → observe loop. This does.

• Accessibility-tree-first perception. The agent reasons over labeled UI elements (describe_ui) and taps by label, not by guessing pixel coordinates — far more robust to layout changes.

• Screenshot fallback. For custom-drawn views that don't expose accessibility, screenshot gives a vision fallback and a way to verify state.

• Loud failures. A tap on a missing label returns the nearest labels on screen, not a silent no-op.

How it works

Agent (Claude / any MCP client)
   goal: "log a leg workout, confirm it appears in History"
        observe → decide → act → observe  (loop)
        │  MCP (stdio)
   ios-agent-driver
        │                         │
   xcrun simctl              idb (+ companion)
   lifecycle, screenshots    accessibility tree,
   deeplinks, permissions    tap / type / swipe by element

Related MCP server: iOS Automation MCP Server

Requirements

• macOS with Xcode (provides xcrun simctl)

• idb for UI perception and actions:

  brew install idb-companion
  pip3 install fb-idb
  idb list-targets   # confirm it sees your booted sim

  Lifecycle tools work without idb; describe_ui / tap / type_text / swipe require it and will tell you how to install it if it's missing.

• Node.js ≥ 18

Install

git clone https://github.com/CodeJonesW/ios-agent-driver.git
cd ios-agent-driver
npm install      # builds via the prepare script

Register with Claude Code

Add to your MCP config (user-level ~/.claude.json, or a project .mcp.json):

{
  "mcpServers": {
    "ios-agent-driver": {
      "command": "node",
      "args": ["/absolute/path/to/ios-agent-driver/dist/server.js"]
    }
  }
}

Or with the Claude Code CLI:

claude mcp add ios-agent-driver -- node /absolute/path/to/ios-agent-driver/dist/server.js

Tools

The loop, by example

A typical agent goal runs as a bounded loop:

GOAL: "open Settings and confirm Notifications is enabled"
1. boot_sim
2. launch { bundle_id: "com.apple.Preferences" }
3. describe_ui            → see "Notifications" cell
4. tap { label: "Notifications" }
5. describe_ui            → assert the toggle state
   (re-read after each action; stop when the goal predicate holds
    or a step budget is exhausted)

The agent owns the loop and the success predicate; this server provides the primitives. That keeps the tool simple and the test logic where it belongs.

Development

npm run build     # compile TypeScript → dist/
npm start         # run the server on stdio

License

MIT © Will Jones (CodeJonesW)