Verity

Preview — Verity is under active development and not yet ready for general use.

Verity is an end-to-end testing tool that combines device automation with LLM reasoning. Write human-readable journey files, and Verity executes them against real devices — navigating UIs, pressing buttons, and verifying what's on screen.

• Android TV
• Android Mobile
• iOS

Journeys are YAML files that describe what a user does and what the app should show:

name: Launch and browse
app: com.example.tv
platform: android_tv

steps:
  - action: "Launch the app"
  - action: "Navigate down to the 'Continue Watching' row"
  - assert: "A 'Continue Watching' row is visible with at least one item"
  - action: "Select the first item"
  - assert: "The detail screen shows a title and a 'Play' button"

Verity splits each journey into segments, hands each segment to an isolated LLM agent, and drives the device through the Maestro SDK. Assertions use a cost-aware strategy: deterministic checks first, then text-based reasoning, then vision — so you pay only for what each assertion requires.

• CLI (verity run) — Run journey files autonomously against connected devices.
• MCP server (verity mcp) — Expose device control as MCP tools for interactive AI workflows (e.g., in Claude Code).

core  ←  device  ←  agent  ←  cli
              ↑                  |
             mcp  ←─────────────┘

• Fast-path key mapping — Common navigation actions (D-pad, gestures) go straight to the device, bypassing LLM generation entirely.
• Subagent isolation — Each journey segment runs in its own LLM session, keeping context windows small and preventing hallucination from long histories.
• Persistent connections — Embedded Maestro SDK holds persistent gRPC/HTTP connections to devices instead of spawning processes per operation.

Minimal path to verify real-device integration:

1. Start MCP server

   ./gradlew :verity:cli:run --args="mcp --transport stdio"
   # or: --args="mcp --transport http --port 8080"

2. Open session — call open_session with platform: android-tv (or android/ios)
3. Press a key — call press_key with key: DPAD_DOWN
4. Check visibility — call check_visible with text: <visible UI text>
5. Capture hierarchy — call capture_hierarchy with filter: content and verify the tree renders
6. Close session — call close_session and verify device state is restored