Start
Phone Use
The mobile runtime for AI agents.
What Phone Use does
Phone Use gives an AI agent an iPhone it can observe, control, and learn. It combines a typed device SDK with an agent-facing CLI, MCP server, outcome verification, reusable app maps, and compiled skills.
| Surface | Best for | Model required |
|---|---|---|
| CLI verbs | Coding agents and shell automation | No |
| MCP server | Tool-capable agent clients | The client supplies the agent |
| phone-use task | A complete natural-language goal | Yes |
| TypeScript SDK | Product and infrastructure integration | No |
| Compiled skills | Reliable repeated workflows | No |
Semantic control
Phone Use reads the native accessibility hierarchy and targets elements by meaning. Vision and coordinates remain available for interfaces that do not expose enough structure.
Checked outcomes
Actions report whether the screen changed, and workflows can assert postconditions. A run completes because the desired state was observed—not because a model said it was done.
Learned routes and skills
App Cartographer turns explored screens into a persistent graph. Known routes and successful workflows replay deterministically with zero model calls.