Test your app through AI vision. Replay 100× cheaper.
Claude MCP AI test automation via Computer Vision. Describe what to test in chat — Claude watches your app, runs the test. E2Easy records it. Every replay costs 1/100th (same test, no vision needed). Full history, analytics, reports. Self-healing when selectors break.
Works with Claude's Computer VisionClaude DesktopClaude.ai (web)Any MCP-compatible clientOAuth 2.1 secured
Why AI vision testing
Three things change when you test through Claude.
Test with AI, pay once
Claude watches once. Replays cost almost nothing.
First run uses Claude's Computer Vision (~$0.12). E2Easy records the test. Every replay: just compute, no vision (recorded interaction stream). Same test, 100× cheaper, forever.
Annual cost · 1 test · 100 replays/year
Claude Computer Use (all runs)
$320
Hand-written Playwright
~$48
E2Easy with Claude vision
$0.12 + $1.80
E2Easy: $0.12 first vision run + $0.018/replay for compute. Computer Use: ~$3.20 per replay (LLM tokens on every step). Playwright hand-written: engineer time ~$48/test/year in maintenance.
Sees the app, not code
Claude watches your UI.
Computer Vision sees buttons, forms, text exactly as users do. No "find this element by id" guessing — Claude understands the visual intent. When UI changes, self-healing fixes visually-broken steps.
// Claude sees the screen
{
"element": "Login button, blue, bottom-right of form",
"action": "Click",
"intent": "Submit login form",
"result": "Navigated to dashboard",
}
Full visibility
Every run recorded. Analytics built in.
Full test history, video on every replay, step logs, performance trends. Analytics dashboard shows flakiness, regressions, and self-heal frequency. Replay cost = visible on invoice.
Sign-up flowPassing
Checkout — guestPassing
Password resetFlaky 82%
Admin invitePassing
How it works
From chat to 100× cheaper replays.
One vision run with Claude. Unlimited cheaper replays. Full history and analytics.
1. Connect Claude
OAuth link your Claude account (Desktop, Claude.ai, or any MCP client). Revocable anytime.
2 min setup
2. Describe what to test
"Test sign-up flow as admin, land on dashboard, check welcome banner." Plain English.
No code, no selectors
3. Claude watches & tests
Computer Vision sees your app, performs actions, asserts outcomes. E2Easy records the video + interaction stream.
Full test recorded · ~$0.12
4. Replay infinitely cheap
Run the recorded test locally, in CI, or our cloud. No Claude vision needed — just replay the interaction stream. Full video & logs.
~$0.018 per replay
Real scenarios
What Claude can test through Computer Vision.
Six examples from real teams. Each starts with a single chat message about what to test.
Visual flow testing
Test the entire onboarding wizard as a new tenant. Fill in form fields, click buttons, confirm landing on dashboard.
Claude sees each screen, understands intent, acts. 17 steps recorded with video.
Claude vision run
$0.15
Per replay
$0.018
Self-heal broken steps
My checkout test is flaky. Identify what's visually broken and fix it.
Claude runs the test, sees which step fails visually, adapts to find the element and recover.
Claude vision diagnosis
$0.02
Fixed in
30s
Cross-browser testing
Test the checkout flow on Safari, Chrome, and Firefox. Tell me where it breaks visually.
Claude runs the same test on 3 browsers, reports visual regressions in each.
Vision runs
3 × $0.12
Each replay
$0.054
Flakiness analysis
Analyze all checkout tests this week. Show me which ones are flaky and why.
Claude reviews run history, video logs, identifies visual and timing issues.
Analysis cost
$0.03
Engineer time saved
~2 hr
Batch replay & analyze
Run all 12 checkout tests against staging. Tell me which ones broke or got slower.
Parallel replay of recorded tests, compare against last run, summarize regressions.
Compute · 12 replays
$0.22
Analysis · vision
$0.00
No-code testing
Test the new pricing page. Walk through sign-up, check all tiers display correctly, verify calculations.
PMs and QA test directly. Claude watches, understands intent, reports visually. No code, no selectors.
Vision run
$0.10
Coding
None
// the math
Vision-based testing costs. Replays, nearly free.
Drop your numbers — test count, annual replays per test, engineer rate. The calculator shows total annual cost: Claude vision + recorded replay compute.
First vision run is smart. Every replay after that is cheap compute — no AI costs.
Approach
LLM cost / yr
Eng. time / yr
Total / yr
Claude Computer Use (vision each replay)
$0
0 hr
$0
Hand-written Playwright/Cypress
$0
0 hr
$0
E2Easy AI vision + replay
$0
~0 hr
$0
You save $0 a year vs Computer Use, $0 vs hand-written Cypress.
Computer Use cost = LLM tokens at ~$3.20 per replay (every run). E2Easy: $0.10-0.15 per vision run + $0.018-0.022 per recorded replay (compute only). Hand-written = engineer maintenance time (~$48/test/year at $80/hr).
The toolset Claude gets
Vision-native MCP API. No selectors. See like Claude does.
10 verbs for vision-based testing. Claude watches, understands, tests, records.
watch_screen(url)→ Screenshot// capture current screen · Claude sees it
click_element(description)→ Result// Claude describes what to click · finds & clicks visually
type_input(text, element_description)→ Result// fill form field by visual description
assert_visible(description, screenshot?)→ bool// check if element/text is on screen
record_test(name, steps[])→ Test// save recorded vision test
list_tests(filter?)→ Test[]// search recorded tests
get_test(test_id)→ Test// get test + step video
replay_test(test_id, env?)→ RunId// replay recorded test · no vision needed
self_heal(run_id)→ bool// Claude fixes broken step visually
list_test_runs(test_id?)→ Run[]// run history · each with video
OAuth 2.1 + PKCE. Short-lived tokens. Revocable.
Destructive verbs require Claude to confirm first.
Side by side
Vision testing, done right. Record once. Replay cheap, forever.
Claude's intelligence on first run. Determinism and cost savings on every replay after.
Claude Computer Use
Vision on every run
Flexible, can adapt mid-test
~$3.20 per 100-step test
Slow: vision + reasoning per step
Probabilistic per run
Hand-written Playwright/Cypress
Engineer-authored
Deterministic, fast
4-6 hours per test to write
Manual fixes on every UI refactor
Selector brittle
Best of bothE2Easy — Vision + Recording
Claude watches once, replays cheap
Claude's vision intelligence · once
Recorded replay · deterministic, fast
$0.12 vision + $0.018 per replay
Self-healing on visual regression
Full history, video, analytics
Case study · TechPay
Vision tests for critical flows. Cost stays flat, coverage grows.
How TechPay scaled E2E coverage using Claude's Computer Vision — recording tests once, replaying cheaply 100× per month, never breaking the LLM budget.
Critical flows
18 vision-recorded tests
Replays / month
~1,800 (daily + pre-release)
Before
Hand-written Playwright + manual fixes
With E2Easy
Claude vision once, deterministic replays
The problem
TechPay's payment flows are mission-critical. Writing and maintaining hand-written tests took 6+ weeks per flow, then every UI refactor needed 3-4 days of selector hunting. They needed intelligence without the per-run cost — Computer Use at $3.20/run × 1,800/month = $5,760/month was unaffordable.
The approach
Switched to Claude's Computer Vision via E2Easy. Opened Claude Desktop, described each payment flow to test ("Walk through credit card checkout, verify receipt."). Claude watched the live app, understood the intent, executed the test. E2Easy recorded everything. Replays became cheap, deterministic, and logged.
The numbers · 4 months in
18 flowsrecorded via vision · in 4 weeks (vs. 6+ weeks hand-written)
$32.40monthly LLM spend for vision + self-healing (Computer Use: $5,760)
100%runs videoed, timestamped, searchable by outcome
We describe what we want to test. Claude watches it happen. Then we replay the same test 100 times for the cost of a lunch. The video library alone is priceless — we can replay any transaction, understand exactly what broke.
Anthropic's Model Context Protocol — the open standard for plugging LLMs into real tools. E2Easy's MCP server gives Claude native verbs for recording, inspecting and replaying browser tests.
Does Claude run my tests every time?
No. Claude authors once. Playwright replays from the saved snapshot — locally, in CI, or in our cloud. Self-healing fires only on the broken step.
Is this the same as Claude Computer Use?
No. Computer Use calls the LLM on every step. E2Easy uses Claude to author the test once from the live DOM, then replays deterministically from snapshots — far fewer tokens per replay, with full video reports.
Do I need an Anthropic API key?
No. Bring your own Claude account (Desktop, Claude.ai, Cursor — anything MCP-compatible). OAuth-based, no static keys.
Will my tests still work without Claude?
Yes. Tests live on Playwright. Claude is the author, not a runtime dependency.
Can other MCP-compatible clients use this?
Yes. The E2Easy MCP server speaks standard MCP over OAuth. Claude Desktop, Cursor, Continue.dev, Cline and similar clients all connect with the same endpoint.
How does Claude know which selectors to use?
It doesn't guess. inspect_page and find_element visit the page in a real Playwright browser and return stable selectors, ARIA, xpath, locators — straight from the DOM.
How is this different from Microsoft Playwright MCP?
playwright-mcp is a raw browser bridge — every replay calls the LLM on each step. E2Easy stores tests as Playwright-grade recordings: replay is deterministic with run history and video reports, and self-healing fires only on the failing step.
Can I use this to fix broken Cypress selectors with Claude?
Yes. CSS, ARIA, xpath, role-based locators — all supported. After a UI refactor Claude reads the live DOM via MCP and produces a drop-in selector fix. Only the failing step is patched.
Does this count as AI test automation or self-healing test automation?
Both. AI test automation for authoring, self-healing test automation for maintenance. Claude patches only the failing step.
What about security?
OAuth 2.1 + PKCE. Short-lived tokens. Revocable any time. Test data stays in your workspace.
Claude watches your app once. Replay it cheap, forever.
Two minutes to connect. Vision testing with full history, analytics, and self-healing. Run tests 100× cheaper after the first vision pass.
