Completed experiment · tool catalogue

79 tools, each with a verdict.

A dated map of practical browser-agent and iOS-agent testing choices. Every row has either local execution evidence or a concrete fit boundary; there are no placeholder “not run” rows.

Catalogue reviewed
20 September 2026
Last experiment
20 September 2026

Coverage is exhaustive within the seven named categories as of the review date, using maintained official projects and materially relevant reference tools. It is a dated map, not a claim that every testing product in existence belongs here.

Evidence key

benchmarked
Repeated locally against the shared Vaultwealth journeys or faults.
screened
A bounded local probe was completed, but the tool was not fully qualified.
setup-blocked
A dated setup or run was attempted, but a named prerequisite or bounded failure prevented a fair result.
source reviewed
Official sources were reviewed and a concrete fit boundary was recorded; no runtime ranking is claimed.

Saved web automation 12

ToolEvidenceCostVersion or boundaryWhat we know
Playwright benchmarked Free / local 1.62.1 Qualified web baseline; fixed assertions and visual comparisons.
agent-browser benchmarked Free / local 0.38.1 Batch reduced its own login overhead; batch edit failed 0/5.
Playwright CLI screened Free / local 0.1.21 Login and edit passed 5/5 but were slower than the saved baseline.
Puppeteer Core benchmarked Free / local 24.43.1 Search passed 5/5 at 2.081 s workflow median and detected stale search; fastest added screen, but not a 2x win.
Selenium WebDriver benchmarked Free / local 4.49.0 / ChromeDriver 151.0.5 Search passed 5/5 at 3.633 s workflow median and detected stale search.
Cypress setup-blocked Free core / paid cloud 16.1.0 package Package pinned; the isolated experiment had no Cypress application binary. No tool-quality claim.
WebdriverIO benchmarked Free / local 9.31.9 Search passed 5/5 at 2.958 s workflow median and detected stale search.
Nightwatch benchmarked Free / local 3.16.0 Search passed 5/5 at 5.425 s workflow median and detected stale search.
TestCafe setup-blocked Free / local 3.7.6 Proxy run did not return a test result before the bounded 60 second timeout on this service-worker fixture.
Taiko benchmarked Free / local 1.5.0 Search passed 5/5 at 23.850 s workflow median and detected stale search; isolated launches were required.
CodeceptJS source reviewed Free / local 4.1.0 package inspected Installed and pinned, but it delegates to Playwright/WebDriver/Puppeteer; no separate driver speed claim after those engines were measured directly.
Testplane source reviewed Free / local WebDriver runner alternative Applicable WebDriver test runner, but it does not change the underlying driver result measured with Selenium and WebdriverIO.

Adaptive browser agents 22

ToolEvidenceCostVersion or boundaryWhat we know
Jev Ultrafast benchmarked Cheap paid selector + local text model 0.1.0 @ 1231850a0bf1a0c0341fe408ef1668dbbfdfac46 Authenticated search passed 5/5; password fields are excluded upstream.
Stagehand screened Free code / model-dependent 3.4.0 Cached replay reached 2.201 s median, 4/4; authoring took 25.544 s.
Browser Use screened Free code / model-dependent 0.13.10 Worked with local Bonsai but passed 4/5 at 90.730 s among successes.
Chrome DevTools MCP screened Free / local 1.9.0 Fast diagnostic companion; not a journey executor.
Playwright MCP source reviewed Free / local 0.0.82 package inspected Pinned locally; an agent transport over Playwright, not a saved-executor speed replacement.
Codex coding agent benchmarked Subscription usage 0.155.1 / gpt-6-astra Stepwise login passed 5/5; the batched instruction arm qualified 0/5.
Skyvern source reviewed Open source / model-dependent Model or agent orchestration layer Computer-vision and LLM browser workflow platform.
LaVague source reviewed Open source / model-dependent Model or agent orchestration layer Web-agent framework with browser actions and retrieval.
AgentQL source reviewed Free tier / paid service Model or agent orchestration layer Natural-language element and page-data queries.
Magnitude source reviewed Open source / vision-model dependent Model or agent orchestration layer Vision-first browser agent; text-only Bonsai is not a fair model.
Browser MCP source reviewed Free / local browser Personal-profile boundary Controls an existing browser through an extension; excluded because the experiment forbids personal browser profiles.
OpenAI computer use source reviewed Paid API Model or agent orchestration layer General screenshot-and-action computer-use model, not a test oracle.
Anthropic computer use source reviewed Paid API Model or agent orchestration layer General computer-use tool requiring an external execution environment.
Browserable source reviewed Open source / provider-dependent Agent framework Self-hostable agent framework; model/orchestration choice rather than a fixed local driver benchmark.
Drisp Browser MCP source reviewed Free / local MCP interaction layer Local browser-control MCP with snapshots and actions; use as an agent interface, not a correctness oracle.
Midscene.js source reviewed Open source / model-dependent Vision model required AI-driven web and Android automation; the selected Bonsai text model is not a fair vision backend.
TestZeus Hercules source reviewed Open source / model-dependent Agent framework Agentic browser testing framework; requires model configuration and independent outcome verification.
Shortest source reviewed Open source / model-dependent Model-dependent test runner Natural-language Playwright tests; model reasoning remains the dominant variable for a fair comparison.
UI-TARS Desktop source reviewed Open source / local or hosted model Desktop GUI agent General desktop GUI agent with browser control; broader than the local web-driver question and still needs an external oracle.
Agent TARS source reviewed Open source / model-dependent General multimodal agent General multimodal agent framework, not a deterministic test executor.
Harness source reviewed Free / model-dependent Experimental agent harness Experimental browser and mobile agent harness; useful for exploration research, without a qualified Vaultwealth run.
Playwright Test Agents source reviewed Free code / model-dependent Authoring and healing layer Planner, generator and healer agents author Playwright tests; repeat execution remains Playwright.

Agent evaluation suites 5

ToolEvidenceCostVersion or boundaryWhat we know
BrowserGym source reviewed Free / model-dependent Evaluation suite, not driver Unifies reproducible web-agent environments; it evaluates agents rather than replacing product test execution.
AgentLab source reviewed Free / model-dependent Research harness Reproducible web-agent experimentation on BrowserGym; not an application regression runner.
WebArena source reviewed Free / local infrastructure Benchmark environment Self-hosted realistic web tasks for agent research; not a substitute for product-specific assertions.
VisualWebArena source reviewed Free / local infrastructure Multimodal benchmark Visual web-agent benchmark; measures general capability rather than Vaultwealth defect detection.
WorkArena source reviewed Free code / ServiceNow instance Enterprise benchmark ServiceNow task benchmark with an application-specific infrastructure prerequisite.

Remote browser infrastructure 5

ToolEvidenceCostVersion or boundaryWhat we know
Browserbase source reviewed Paid service Remote infrastructure outside local-only scope Managed remote browsers; Stagehand integration is first-party.
Browserless source reviewed Self-hosted or paid Remote infrastructure outside local-only scope Remote Playwright/Puppeteer browser infrastructure.
Steel source reviewed Open source or paid Remote infrastructure outside local-only scope Open-source browser API for agents and apps.
Hyperbrowser source reviewed Paid service Remote infrastructure outside local-only scope Managed browsers and agent infrastructure.
Anchor Browser source reviewed Paid service Remote infrastructure outside local-only scope Cloud browsers for agents and automation.

Visual verification 11

ToolEvidenceCostVersion or boundaryWhat we know
Vitest browser visual assertions source reviewed Free / local Visual oracle or diff layer Screenshot matching inside Vitest browser mode.
Percy source reviewed Paid service Visual oracle or diff layer Managed cross-browser visual review.
Applitools Eyes source reviewed Paid service Visual oracle or diff layer Managed visual-AI comparison and review.
Chromatic source reviewed Free tier / paid service Visual oracle or diff layer Storybook-oriented component visual testing.
Argos source reviewed Open source / paid service Visual oracle or diff layer Visual and accessibility-tree diffs for pull requests.
Lost Pixel source reviewed Open source Visual oracle or diff layer Sunsetting; useful comparison history, not a fresh adoption recommendation.
BackstopJS source reviewed Free / local Visual oracle or diff layer Self-hosted browser screenshot regression testing.
Loki source reviewed Free / local Visual oracle or diff layer Storybook screenshot regression testing; verify maintenance before adopting.
reg-suit source reviewed Free / local or object storage Screenshot diff pipeline Screenshot regression workflow; requires a capture source and does not explore the application.
TestivAI OSS source reviewed Open source / model-dependent AI visual test framework Open-source visual testing framework; model and review configuration are additional variables.
Odiff source reviewed Free / local Image diff primitive Fast image comparison primitive, not a journey runner or visual-understanding agent.

iOS and mobile 20

ToolEvidenceCostVersion or boundaryWhat we know
Maestro benchmarked Free CLI / optional cloud 2.6.1 / Java 21 Retained iOS saved-flow baseline.
AXe benchmarked Free / local 1.8.0 Fast warm runs, but cold accessibility and text-entry reliability blocked adoption.
XcodeBuildMCP screened Free / local 2.7.0 Useful installed-app probe; full save/relaunch journey was not qualified.
XCUITest setup-blocked Included with Xcode Native interaction or simulator layer Vaultwealth checkout lacked an app-owned Xcode UI-test target.
Detox setup-blocked Free / local Native interaction or simulator layer Vaultwealth checkout lacked Detox configuration and an app-owned target.
Appium source reviewed Free / local Native interaction or simulator layer Cross-platform W3C WebDriver automation; iOS uses the XCUITest driver.
Appium MCP source reviewed Free / local Native interaction or simulator layer Official MCP layer for Appium mobile automation.
Patrol source reviewed Free core / optional cloud Native interaction or simulator layer Flutter-focused native integration testing.
EarlGrey 2 source reviewed Free / local Native interaction or simulator layer Google iOS UI automation integrated with XCUITest.
idb screened Free / local fb-idb 1.1.7 Isolated runtime found the simulator and returned its accessibility tree through ios-simulator-mcp; the broken user-level install was not modified.
idb-mcp source reviewed Free / local Native interaction or simulator layer MCP wrapper around idb for iOS Simulator control.
AppleSimulatorUtils source reviewed Free / local Simulator utility, not journey runner Simulator state and permission utility; upstream deprecates overlapping operations in favor of simctl.
SimPilot source reviewed Free / local macOS simulator agent Agent-oriented iOS Simulator automation; requires a separate product-state verifier for defect claims.
Wand source reviewed Free / local Private-framework risk Fast simulator automation built around private Apple frameworks; unsuitable as the default long-lived baseline without accepting that maintenance risk.
ios-simulator-mcp screened Free / local 2.1.0 Started with 17 tools, found the dedicated simulator and returned a 6,940-character accessibility description using isolated idb.
ios-simulator-mcp by Yael Gilboa source reviewed Free / local Alternative MCP fork Alternative simulator MCP focused on screenshots, taps and accessibility; overlaps XcodeBuildMCP and AXe.
iosef source reviewed Free / local Accessibility automation CLI Accessibility-first iOS Simulator control; a candidate for a future native screen, but no claim is made beyond interface fit.
Mobile MCP screened Free / local 1.0.4 / device agent 0.0.26 Started with 32 tools, found the simulator and listed 27 elements; its temporary device agent was removed after the probe.
AutoMobile source reviewed Free / model-dependent Mobile exploration agent Agentic mobile exploration project; model cost and nondeterminism require separate accounting.
Mobilewright source reviewed Free / local 0.0.57 package inspected Installed transitively with Mobile MCP; early project, so baseline maturity and API stability remain boundaries.

Local model helpers 4

ToolEvidenceCostVersion or boundaryWhat we know
Ternary Bonsai 2 27B PTQ1_0 screened Free / local revision 6ed5e12bf84b7a63069882c91dd9e9218647d17b Text helper passed 25/25; no-thinking median 1.082 s at 7.305 GB sampled peak RSS.
Prism llama.cpp screened Free / local 0.2.0-dev build 10709 @ 9a9394a895b96003ca842a6041cb28ac49a108f7 Pinned runtime used to serve Bonsai locally.
Ollama source reviewed Free / local Model runtime, not UI executor Convenient local model runtime; not benchmarked in this experiment.
MLX LM source reviewed Free / local Model runtime, not UI executor Apple-silicon local model runtime; not benchmarked here.