Testa MCP server

Testa

Drive the iOS Simulator like a human — from an AI agent.

Real HID touches (every gesture), screen reading via the accessibility tree or on-device OCR, token-efficient, fast, and zero third-party dependencies. Test React Native / Expo and native SwiftUI apps end-to-end — without adding a single testID — then save the run as a flow file and let CI replay it for free.

An agent tapping, pinching, rotating, dragging-and-dropping, and typing — all verified through the accessibility tree, no screenshots.

Why Testa

Agents are great at writing iOS apps and clumsy at the part that comes next: actually exercising them in the simulator. Testa is built for that — and three things set it apart from the rest of the field: it drives screens that expose zero accessibility (via on-device OCR), it turns an agent session into a deterministic flow file that CI replays with no model in the loop, and it's a single, fully-open, dependency-free binary.

Apple Vision OCR to tap any visible text — so it drives canvas, games, WebViews and vibe-coded apps that never added a testID.

• 🧠 No app setup required. Reads the accessibility tree, and falls back to

— a ui snapshot of a real screen measures ~203 tokens against ~1,500 for the same screen as an image. Measure it yourself →

• 🪙 Token-efficient. One compact line per element (e5 Button "Save" #save @120,300)

reply ends with -- ui changes --, so the loop is act → read the diff, not act → full snapshot.

• 🔁 Acting commands answer with the UI diff. Every tap/type/swipe

what the agent just did as plain text; testa flow run smoke.flow replays it deterministically — zero tokens, JUnit output, failure artifact bundles.

• 🎬 Record once, replay forever. testa flow record save smoke.flow writes

client hot: ~60 ms per snapshot (measured 24–39 ms marginal, see bench/).

• ⚡ Fast. A warm daemon keeps the connection, accessibility translator and HID

pinch/zoom, rotate, multi-touch, hardware buttons, key combos, unicode/emoji text — genuine HID events.

• 👆 Every gesture, for real. Tap, long-press, swipe, drag-and-drop,

Face ID / Touch ID, dark mode, Dynamic Type, status bar, locale, pasteboard, photo library, launch env & args.

• 🌍 Control the device, not just the app. Push notifications, GPS location,

targets, duplicate labels) that fails a CI job.

• ♿ testa audit — accessibility audit (missing labels, sub-44 pt tap

diff, a red heatmap, and OCR-aware - lost: / + new: lines so a percentage becomes a sentence.

• 🖼 testa vdiff — visual regression with an antialiasing-tolerant pixel

a Claude Code plugin manifest, and a composite GitHub Action.

• 🔌 Agent-native. Ships an MCP server (testa mcp), a Claude Code skill,

SimulatorKit, AccessibilityPlatformTranslation, Vision and simctl.

• 🔒 Local & private. A 0600 per-user Unix socket. No network, no telemetry.
• 📦 Zero third-party runtime deps. Talks straight to Apple's CoreSimulator,

Shortened. The full README is on GitHub.