iOS Simulator E2E driver for AI agents. Autonomous testing and interaction with iOS simulators via MCP protocol.
Testa provides 54 tools with consistent naming (verb-first: ui, see, find, tap, etc.) and descriptions present for all tools. However, descriptions are terse (avg ~40 chars, well below the 194-char baseline for A+ tools), lacking context on WHEN to use each tool or dependencies between them. Input schemas are present and mostly typed, but many lack depth: parameters like 'selector' in find/tap/assert are under-documented (what formats are valid? eN vs #id vs "label"?). Output schemas are not documented, agents cannot predict what fields to expect. Error handling is absent from descriptions. Security considerations (e.g., permission gates for destructive tools like terminate, shutdown) are not mentioned. Composition is strong (tools are single-purpose), but chaining guidance is missing.
add photos/videos to library
set appearance (dark or light)
list installed apps
assert element state (exists|gone|value=..|label=..)
accessibility audit (labels, 44pt targets, dupes)
enroll/unenroll/match/nomatch biometry
boot simulator by UDID or name
press hardware button
clear element value
Descriptions are uniformly terse (avg ~40 chars vs 194-char baseline). Most lack context on WHEN to use the tool, dependencies, or prerequisites. E.g., 'tap element or coordinates' does not explain selector format (eN | #id | "label"), fallback behavior, or when to use tapocr instead.
Output schemas are not documented. Agents cannot predict what fields to expect from tools like 'see' (OCR output), 'find' (element list), 'logs' (log format), or 'crashes' (crash report structure). This forces agents to guess or make follow-up discovery calls.
Parameter descriptions lack format/constraint details. 'selector' in find/tap/assert is described as 'element selector' but valid formats (eN, #id, "label") are only in the CLI usage text, not in the schema description. LLMs cannot reliably infer these constraints.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 49 | 2026-07-28+ | v2 |
set content size
get newest crash report
list booted simulators
drag gesture
drag and drop gesture
elements matching label/id/value/role
run deterministic replay flow
get simulator info
install app on simulator
press hardware key
press key combination
launch app with optional environment and arguments
set simulator locale
set or clear simulator location
get recent app console logs
long press gesture
run flow on multiple devices
open deep link or universal link
copy text to pasteboard
paste text from pasteboard
grant/revoke/reset app permission
pinch gesture
push notification to app
record video
rotate gesture
take a screenshot
scroll until an element is visible
OCR every visible text + tap coordinates (any app)
install Claude Code skill and register MCP server
set element value
shutdown simulator
start daemon for simulator
check daemon status
set or clear status bar
stop daemon
swipe gesture
tap element or coordinates (falls back to OCR text)
tap OCR text
terminate app on simulator
type text
type text into element
on-screen snapshot (diff=changes, full=incl. off-screen)
visual regression (writes baseline on first run; OCR-aware diff after)
get testa version
wait until element appears (or disappears)
Destructive tools (terminate, shutdown, delete-like operations) lack error handling guidance or confirmation patterns. No description warns agents of irreversible consequences or suggests dry-run/confirmation steps.
No tool chaining guidance. Agents cannot infer that 'find' should precede 'tap', or that 'launch' requires a bundle ID from 'apps'. Descriptions do not mention dependencies or suggest multi-step workflows.