Scored across 77 tools
Multiple tools have unclear boundaries. The UI inspection cluster (describe, native-describe-screen, native-full-hierarchy, native-find-views, native-view-at-point, native-user-interactable-view-at-point, debugger-component-tree) has several close cousins, especially the two point-inspection tools and native-describe-screen vs describe. The typing cluster (keyboard, text, key, paste), the two network-log tools (view-network-logs vs native-network-logs), the CPU profiler queries (react-profiler-cpu-summary vs profiler-cpu-query), and the rotate vs gesture-rotate naming collision further compound the ambiguity, though the descriptions are detailed.
Naming is consistent within clusters — gesture-*, native-*, debugger-*, flow-*, chromium-* — but the overall set mixes conventions: verb_noun (list-devices, boot-device), bare nouns (button, keyboard, text, key, describe), and inconsistent family naming (profiler-cpu-query vs react-profiler-cpu-summary). It remains readable, but there are notable deviations like debugger-component-tree and screen-recording-start vs screenshot.
77 tools far exceeds the well-scoped range and is too many for the apparent scope, even though the server spans a broad platform (device automation, profiling, network, flows). The count is inflated by near-duplicate tools (multiple UI inspectors, CPU queries, and the keyboard/text/key split) and meta tools (update-argent, dismiss-update, gather-workspace-data). An agent faces an unwieldy surface to navigate.
The tool surface is remarkably comprehensive: device lifecycle, gestures, UI inspection, network capture, JS/native debuggers, React and native profilers, screen recording, permissions, and flow recording/replay are all covered. Minor gaps exist — no location simulation, biometrics, or device file push/pull — but agents can work around these, and core automation workflows have no dead ends.