Opposable, a new tool covered by The Register, aims to give AI agents 'thumbs' by letting them operate a smartphone through screen taps, swipes, and typing. The Android version reads the screen via accessibility services and screenshots, with a vision-language model deciding where to tap. iPhones require a companion computer, and about 40 common tasks run as fixed code rather than through the model.
The company's own demonstrations show the approach is not quick. Changing a phone to dark mode and setting a five-minute timer took 50 seconds and 17 tool calls. Looking up Apollo 11 launch details took nearly two minutes and cost $0.44. An on-device example, opening the clock app and switching to the stopwatch tab, took about 96 seconds.
Reliability is also questionable. On eight tasks the model had never been tuned on, run twice each on an emulator, it succeeded 4 and 3 times. When given one subgoal at a time, it succeeded 6 times in both rounds. The tool requires sideloading an APK and granting access to read and interact with the screen, and the company warns that whoever holds the token can operate the phone.
The Register notes that Opposable is marketed as a way to 'hand off phone chores' rather than as an accessibility aid, and asks why anyone would use it given the speed, cost, and failure rate. The examples suggest the main benefit is avoiding the use of one's own hands, at the price of tokens and patience.