block/trailblaze

What it solves

Trailblaze provides a way for AI coding agents to control real devices (iOS, Android, and web) using natural language. It bridges the gap between an agent's ability to write code and its ability to interact with a live application UI, allowing developers to use agents to drive apps, record those interactions as deterministic tests, and maintain them through AI-powered self-healing.

How it works

Trailblaze provides a CLI that coding agents (like Claude Code, Cursor, or Goose) can use to interact with devices. The agent uses snapshot to read the UI tree and tool to act on specific elements. These sessions are saved as "trails"—YAML files that record both the natural-language intent and the platform-specific actions. These trails can then be replayed deterministically in CI without an LLM, or re-run with a --self-heal flag that uses a built-in agent to patch small UI changes.

Who it’s for

It is designed for developers and QA engineers who use AI coding agents to automate the end-to-end testing of mobile and web applications across different platforms.

Highlights

• Cross-Platform Control: Supports iOS, Android, and web with native fidelity, exposing the full accessibility tree or DOM rather than a limited intersection of features.
• Agent-Ready Skills: Ships with a portable SKILL.md format that teaches agents how to use the CLI, ensuring the agent knows exactly how to drive the device.
• Deterministic Replay: Converts AI-driven sessions into committed YAML trails that run in CI without requiring an LLM at runtime.
• AI-Powered Self-Healing: Automatically patches minor UI drift (like text changes or popups) during replay to prevent brittle test failures.
• Custom Tooling: Allows developers to define typed custom commands (e.g., login or addToCart) in TypeScript for more complex agent interactions.
• Trace Viewer: A browser-based report viewer providing screenshots, view hierarchies, and video for every step of a run.