Most attempts to automate phones with AI lean on screenshots and a vision model guessing where to tap — slow, costly, and fragile the moment a layout shifts. DroidRun takes the opposite route: it reads the device's accessibility tree and serializes the live UI into structured text an LLM can reason over directly, treating the phone like a queryable document rather than an image to interpret.
What Sets It Apart
- Text-first, vision-optional: the accessibility tree yields precise element identity and bounds, so the agent acts on real UI nodes instead of pixel coordinates; vision mode stays available for cases plain text can't cover.
- Model-agnostic by design: swap between OpenAI, Anthropic, Gemini, DeepSeek, or a local Ollama model without rewriting agent logic — handy for trading cost against capability per task.
- Real device control, multiple entry points: drive automation from a one-off CLI command, a Python workflow, Docker, or an interactive TUI, against physical or emulated Android (and now iOS) hardware.
- Built-in observability: traces export to Arize Phoenix and Langfuse, so you can see why an agent tapped what it tapped — the part that usually breaks down in multi-step mobile flows.
Who It's For
Great fit if you're building QA automation, app-testing pipelines, or agentic workflows that must touch real mobile apps with no public API, and you want LLM choice rather than a locked-in backend. Look elsewhere if you need a no-code consumer tool, guaranteed determinism, or you're only automating a single web app — a browser agent will be simpler and steadier. Reliability still rides on the quality of each app's accessibility data, which varies widely across the store.
