Phonedriver lets AI agents use phones with 6.9 times less token use
The 6.9 times number comes from a test that the maker made, not from other tests.
Phonedriver is an open-source tool. It lets AI agents, such as Claude Code and Codex, use an Android phone or an iPhone. It reads each screen as short text, not as a picture. The maker tells that it used 6.9 times less tokens than the tool mobile-mcp. The test had 9 tasks and was on 10 November 2026. Tokens are the units that use up the limit of an agent.
How it works
The agent gets each button on the screen with a number and taps by number. After each tap, the tool sends only the changes. A small helper on the phone keeps the screen connection open.
The numbers
- 9 tasks, each done 2 times, with Claude Sonnet 5.5: 6.9 times less tokens than mobile-mcp.
- On a HONOR X8 phone, a screen read was 78.5 ms against 2,507 ms with standard adb tools.
- A script read the data of the phone after each task, not the report of the model.
What is not known
The maker tells the makers of the other tools to report an incorrect test, and will do it again. A payment check holds a tap on buttons such as Pay and tells the user to select. It does not find buttons that are only icons, and a user can turn it off.
Insight: Token cost is a new way for agent tools to compete
OpenRouter found that one MCP server (a standard connection between agents and tools) used 59,417 input tokens on a GitHub task, against 1,421 for a command-line tool. Phonedriver gives a token number against a different MCP server. This reads as a race to a lower cost, where the tool with the low cost can be selected. Tell each vendor to give token counts on your tasks.
Sources
Posted
