Verification commands vs live terminal
Choose the correct execution surface for testing or real project processes.
Guides follow the current project source. Check the release notes for your installed version.
Choose a project in Commands
Commands combines the selected project’s live terminal, managed processes, application launcher, browser controls and monitoring. Confirm the project and execution surface before running work.
The app exposes command policy separately from file access and AI-change policy. Read-only access and native sandbox requirements still matter; inspect a denial rather than attempting an unrestricted fallback.
Run disposable verification
Use a discovered build/test/check/lint preset when a temporary copy is sufficient. This path disables networking, bounds execution/output and discards incidental writes after the command.
Supported discovery includes safe-named package scripts, Cargo checks/tests/builds, Python test discovery and Go tests/builds. Arbitrary shell strings, deployment/publish commands and development servers are not disposable presets.
Run a live one-shot command
Use Live terminal for a finite command in the real approved project when the work needs persistent changes or permitted network access. Choose a workspace-relative working directory if the command belongs below the root.
Inspect status, duration, exit code and captured output. A nonzero exit or a queued approval is not a successful command result.
Start and manage a long-running process
Use a managed process for development servers, watchers and background workers. Follow its running/exited/stopped state and incremental logs, then stop or restart the specific process when appropriate.
Managed identity/history survives normal MCP/UI reconnects. Full app-restart reattachment is currently advertised on Linux; Windows/macOS remain unavailable until equivalent native recovery is validated.
Approve through the local controls
The disposable command policy can be review, automatic or disabled. Pending command definitions are rechecked before local approval so a changed script cannot silently replace the one you reviewed.
In AI Auto, supported live commands/process starts can execute immediately. In AI Review they follow the applicable command policy and may wait for local Accept/Reject. MCP cannot approve its own pending request.
Inspect the application and recorded evidence
Use the launcher and managed browser controls for UI verification when the task needs them, then correlate output and monitoring state. Confirm a local preview responds rather than assuming a started server is ready.
Pause AI is the emergency boundary for active AI live execution and managed process groups. Keep the result tied to the exact project and runtime you inspected.