Skip to main content
Mobster gives your agent tools and a skill that describes the testing loop. Ask your agent to run a flow on a simulator. Save the flow as a check. Run the check again on every pull request. For Cursor, Windsurf and Xcode’s agents, follow the setup below from each one’s documentation. Each line needs Mobster for Mac in Applications, opened once. It adds the mobster MCP server and the skill, and changes nothing else. Run it in a terminal, then start a new session in the agent. Restart Cursor. For Cursor’s one-click link and what the install prints, see Cursor. The long path works before mobster is on your PATH. Add --project . to write this repository’s .cursor/mcp.json instead of yours; Xcode’s agents have no project config. Then run mobster mcp doctor: it starts the server the way the agent would and lists its tools. Claude Code uses the plugin instead. Mobster writes the files named in mobster mcp install --dry-run, keeps a .mobster-backup beside each file it changes, and never writes a key.

​
Xcode’s own tools

Xcode’s agents see Xcode’s simulator tools too. Use those to build and look at the app. Use Mobster to check saved assertions, replay a saved check without a model, or test on a real iPhone. mobster mcp install claude-code codex --with-xcode-tools also gives those two agents Xcode’s own tools.

​
The loop to ask your agent for

  1. After you change a screen, ask your agent to build for the simulator. Ask it to call verify_start with the steps and what must be true after them. Then ask it to drive the app through those steps. Finally, ask it to call verify_finish.
  2. When the verdict is passed, ask your agent to call save_check. The file lands in .mobster/checks. Commit it with your change.
  3. Before the pull request, ask your agent to call run_tests with pr: true. Mobster runs the saved checks that cover the changed screens. It drafts a check for each changed screen without coverage.
  4. On every push, mobster ci runs the same checks on a Mac of yours and posts the verdict and each failure’s video on the pull request (CI).
Mobster can replay a passed check’s approved steps. A changed step may need Smart to heal it; the proposal waits in review_heals for your review. Text that Smart needs to write again can also require a model call. Put this in AGENTS.md, CLAUDE.md or .cursor/rules so the agent runs the loop without being asked:
AGENTS.md

​
Crashes and several agents

Give your agent a crash report. Ask it to call repro_crash. The tool writes a check that follows the crash’s path. It runs the check on a simulator with the screen recorded. It reports whether the app crashes the same way. After the fix, ask your agent to run the check again. Each agent’s mobster mcp leases its own simulator, so two agents in two editors drive two simulators. When none is free the call says who holds each. Cursor and VS Code also have a Mobster extension: a side panel with your app live on a simulator, your checks and their approvals (Editor extension). In Codex, Windsurf and Xcode, the skill and the MCP tools above are how an agent uses Mobster.