Your agent might tell you a feature is finished because the code compiles, but you won't know if a button is broken or a sheet fails to close until you open the app yourself. This leads to manual verification and constant back-and-forth to fix things the agent missed. Additionally, every new session starts from zero, wasting tokens and time as the agent re-learns how to build, launch, and navigate your app.
Swift Verify lets your agent write an app-specific verify skill that lives in your project. This skill remembers how to launch the app and reach each feature, so later sessions can get straight to work. Instead of just taking the agent's word, you get visual proof: screenshots, screen text, and video of the agent actually driving the feature in a simulator.
Features
- A persistent, app-specific skill. The agent creates a skill that knows how to build, launch, and navigate your app. Because it lives in your project, subsequent sessions can use it immediately instead of re-discovering your setup.
- A growing map of your features. The agent maintains a file for each feature, recording how to reach it, how to drive it, and any traps it found during previous runs.
- Visual proof you can see. Every step produces a screenshot and screen text. For flows with motion, the agent provides a video, allowing you to review the exact user experience without opening a simulator.
- Clean, isolated simulators. Each project uses its own dedicated simulator copy. This ensures every run starts with no leftover data and keeps your personal simulators untouched.
- No test code in your app. The agent drives your app exactly as it is without adding debug screens or extra labels.
- Support for Mac apps and CLI tools. In Claude Code, the agent can drive Mac apps using computer use, after you approve it. It also handles command-line tools by using terminal output as proof.
How to use it
You can trigger the agent to create or run verifications using slash commands in Claude or prefixing with $ in Codex.
- "
/create-verification-skillverify the sign in flow" - "Verify the new search feature in the running app."
- "Show me the add-book flow on the simulator, with a video."
- "I just finished the settings screen. Check it in the app before I merge."
- "Does the item I add still show up after I relaunch the app? Prove it."
- "Add the new sharing feature to the verify map."
FAQs
Does this replace my UI tests?
No. Use XCUITest for core journeys that must always work. Use this skill to drive new flows and verify visual results quickly.
What can't it prove?
Since it runs in a simulator, it cannot prove behavior related to the camera, push notifications, iCloud, real network conditions, haptics, or the speed of a real device.
Which agents does it work with?
It is built for Claude and Codex. Driving a Mac app requires computer use.
