← All work
SHEET 01 / 04

chauffeur

Lets coding agents drive the iOS Simulator — and tells them the truth.

180-run benchmark vs Xcode 27’s MCP: under half the turns at ~1/6 the cost, 45/45 tasks verified on Opus.

  • Swift
  • MCP
  • XCTest
  • Homebrew

GitHub ↗Watch the benchmark ↗

FIG. 01 — NO EFFECT — the tap changed nothing, and the agent is told so
FIG. 02 — Same task, both tools: 5 turns vs 14
FIG. 03 — APP CRASHED — with the line that caused it

The problem

Coding agents can write an iOS app but can’t tell whether it works. Pointed at the Simulator, an agent taps a disabled button and reports “Done” — nothing happened, and it has no way of knowing.

What I built

chauffeur is an MCP server and agent skill, written in Swift, that lets Claude Code, Codex or Cursor drive the iOS Simulator and get the truth back. It installs with Homebrew and ships a doctor command that checks your Xcode setup.

How it works

It hands the agent the screen as an accessibility outline it can reason about, taps through the Simulator’s own input system, then checks what actually changed. A tap that did nothing returns NO EFFECT, one an alert swallowed returns INTERCEPTED, and a crash returns APP CRASHED with the line that caused it.

Outcome

Benchmarked head to head with Xcode 27’s own MCP server over 180 runs: on Opus it verified 45 of 45 tasks to Xcode’s 39, in under half the turns at roughly a sixth of the cost — and it answered every crash and log task Xcode couldn’t.