Ship without holding your breath.

A deploy can quietly break signup, login, or checkout — and you find out from the customer who already left. Prufa catches what's broken in minutes, after every deploy.

No credit card required · No signup to run a free audit

See what Prufa finds on real sites →
49
Show HN apps audited
38
had a critical finding on launch day
14
side-project launches audited
11
of those sent zero analytics events

Do you find out your signup is broken, or does your customer?

Usually the customer. A deploy can silently break signup, login, or checkout while every page still loads and no error is thrown — so nothing alerts you, and the first signal is a lost sale or a support email. Here are the two ways it happens.

Damage vector 1

A deploy breaks checkout — you find out four days later.

Prufa catches this on the deploy — instead of a customer emailing you four days later.

Damage vector 2

A deploy breaks signup — and new users silently bounce.

No one emails you about a signup they couldn't finish — the page still loaded, so nothing alerts you.

Both of these are preventable. Both are findable in minutes if someone is watching.

How does Prufa work?

Prufa is an AI QA engineer for web products: it runs your signup, login, and checkout flows in a real browser after every deploy and machine-verifies the result. Three steps to your first audit — free, no signup, no card for the first one.

1

Paste your URL

Prufa runs a real-browser audit of your site — tracking tags and consent flows, UX antipatterns, and your key money flows (signup, login, checkout). The verified report lands in your browser in about 60 seconds. No account. No npm install. No Playwright config.

2

Read your findings in two tiers

The report splits into two tiers. Verified findings: deterministic code caught a real breakage — a beacon that didn't fire, a redirect that returned the wrong status code. Named check. Evidence line. Green result. Advisory findings: the LLM flagged something worth a human look — not a machine verdict, labeled clearly, never phrased as "broken." You always know which is which.

3

Watch it after every deploy

Connect your site. Every push to your main branch triggers a re-run of your money-flow checks via a deploy webhook. When a flow that passed before now fails — a Slack alert fires with a screenshot and step trace showing exactly what broke and when. Delta-only: you are never warned twice about the same thing.

Money Flows

Signup, login, and checkout — tested step by step in a real browser, after every deploy.

Prufa drives your signup, login, and checkout flows with a real browser and an LLM-backed agent — then asserts each step with deterministic code: form fields, redirects, HTTP status codes, post-submission beacons.

When a step fails, you get a verdict: the step name, what the assertion expected, what it got, and a screenshot of the browser at the moment of failure. Not an advisory warning. A named, machine-graded result.

The LLM navigates. Plain code verifies.
That is why a Prufa finding is machine-graded evidence — not an opinion about what might be wrong.

Chaos QA

What does your signup form do when someone pastes a 200-character email?

Gremlin mode runs your money flows under adversarial conditions — mutating inputs, timing, and edge-case paths that a standard test never explores. A normal QA run tests the happy path. Gremlin tests what happens when it isn't.

Each Gremlin run produces a "most valuable finding": the specific mutation that caused a flow to fail. The password reset that works — until the email contains a "+" character. The checkout that completes — unless the discount field is filled first. These are real bugs in real code.

Gremlin mode is on the Pro tier ($99/mo). Dry-run mutations by default — no actual data is corrupted.

Audit Report

Every release, witnessed — machine-graded findings your team can share.

Every audit produces a shareable HTML report with a public URL. No login required to open it. Findings split into two tiers: Verified (a deterministic check caught a real breakage — named check, evidence line, timestamp, green badge) and Advisory (worth a human look, clearly labeled, never phrased as "broken"). You always know which tier you're reading.

Every report opens with a plain-language verdict and ranks findings by severity and product impact — auth, signup, and payment breakages surface first — so the thing most likely to cost you a customer is at the top, not buried in a list.

Agent Surface

Can your Claude Code agent verify that signup still works after a deploy?

Yes. The Prufa MCP server is on the Official MCP Registry. Install prufa-mcp in any MCP-compatible host — Claude Code, Cursor, Cline — and your AI coding agent can call prufa.run_audit as a tool call.

The same 3-step monitoring setup that works from the dashboard works from a 5-tool-call agent session. No separate "developer mode." No pared-down API results.

Integration callout CLI · HTTP API (OpenAPI-published, versioned) · MCP server (Pro tier)
Works with: Claude Code · Cursor · Cline · any MCP-compatible host
Listed on: Official MCP Registry · Glama · mcp.so
Also Covered

The same run also checks your tracking — as part of the audit, not the point of it.

Money flows come first. But because Prufa is already driving a real browser, the same run captures every analytics beacon at the network level — so you also find out which conversion pixels stopped firing and which consent flows stopped gating, before your ad platform sends an anomaly report two weeks later. It's one finding type among many that Prufa watches, not the product.

Across the 49 Show HN apps audited, broken or misconfigured tracking was the single most common finding — one signal Prufa surfaces alongside the money-flow checks that lead.

What we found when we audited 49 real launches

"Of 49 Show HN launches audited, 38 had at least one critical finding on the day they shipped. Broken tracking was the most common verified failure — including sites actively running paid ads whose conversion pixels had stopped firing."

Prufa audit study · prufa.dev · cross-posted to dev.to

"Across 14 side-project launches, 11 sent zero analytics events since going live. The pages loaded. The business events — signups, clicks, conversions — were never recorded anywhere."

Prufa audit study · prufa.dev · cross-posted to dev.to
11/14
side-project launches that sent zero analytics events
38/49
Show HN launches with a critical finding on launch day
Every finding includes a named check, a Geist Mono evidence line, and a timestamp.

Does Prufa work with Claude Code, Cursor, and AI coding agents?

Yes. The MCP server is first-class.

prufa-mcp is on the Official MCP Registry. Any MCP-compatible host — Claude Code, Cursor, Cline — can call prufa.run_audit and get back the same JSON report of verified findings a human sees in the dashboard. The agent surface ships from day one: CLI, HTTP API (OpenAPI-published, versioned), and MCP server.

Set up monitoring in a 5-tool-call agent session, or from the dashboard. Same result either way.

Listed on
Official MCP Registry
Glama
mcp.so

Isn't Prufa just a Playwright wrapper?

Playwright is the browser. Prufa is the verification layer that runs after the browser.

A Playwright script records a click sequence and replays it. When Prufa drives a checkout flow, a real browser runs with an LLM-backed agent navigating — then deterministic CDP-level code asserts every beacon, every redirect, and every HTTP status code. The LLM figures out the path; the code grades the result.

Script-generation tools give you more test scripts to maintain. Prufa gives you a continuous QA watchdog — no test author required, no recorded-script rot when your UI changes.

Will Prufa runs show up in my GA4, PostHog, or Segment?

No. Prufa runs in an isolated browser session on a dedicated infrastructure domain. Analytics events from Prufa runs are tagged with $lib=prufa-backend — filter them out of GA4, PostHog, or Segment in one line of each platform's filter config.

Prufa never touches a real payment instrument. Checkout flows are tested up to the payment step, then stopped.

Monitoring never corrupts the data it protects.

Start free. Add monitoring when you're ready.

Run a free audit with no account or card. Paid plans start a 7-day trial — card upfront, auto-bills after day 7, cancel anytime before.

Free
$0
One-shot audits and flows — no signup, no card, free forever.
  • Anonymous one-shot audit — six graded sections, performance to compliance
  • LLM-backed agent loop for one-shot flows
  • 7-day re-check by email — see what changed
  • Shareable report, HTML and JSON
  • Read-only report API access
Audit my site →
No credit card. No account. Results in about 60 seconds.
Starter
$29 /mo
Your core flows — signup, login, checkout — re-verified daily and on every deploy.
  • Plain-English test flows, compiled to reviewable specs and run in a real browser
  • Daily reruns plus deploy-trigger checks on every monitor
  • Slack and email alerts the moment a flow breaks
  • Gremlin chaos mode — capped to 20 steps per run, dry-run by default
  • Hard cap by default with optional overage
Start 7-day trial →
Card upfront · auto-bills after day 7
500 runs/mo · 3 sites · overage $0.10/run, opt-in
Most popular
Pro
$99 /mo
Deeper flows — payments, logged-in journeys — across staging and production, drivable by your agent.
  • Everything in Starter
  • Gremlin chaos mode — capped to 40 steps per run, full mutation authorization
  • Money-flow monitoring and credential-backed flows (encrypted vault)
  • MCP server and agent skill access
  • Staging and production in one workspace, shared quota
Start 7-day trial →
Card upfront · auto-bills after day 7
5,000 runs/mo · 10 sites · overage $0.07/run, opt-in
Team
$179 /mo
Every flow you have, owned by a team — seats, roles, and org-scoped API keys.
  • Everything in Pro
  • Gremlin chaos mode — capped to 60 steps per run, shared across the team quota
  • 5 seats included, then $29/seat
  • Org-scoped API keys and roles
  • Priority support path
Start 7-day trial →
Card upfront · auto-bills after day 7
50,000 runs/mo · unlimited sites · overage $0.05/run, opt-in

Runs meter one LLM call each; confirmed-flow replays are never metered. Full pricing & billing details →

Common questions

Answers in plain English. No marketing speak.

Do I need to set anything up to run an audit?

No. Paste your URL into the field and hit "Audit my site." The report opens in your browser in about 60 seconds. No account, no npm install, no API key, no Playwright config. The audit runs on Prufa's infrastructure.

Continuous monitoring (re-runs after every deploy, Slack alerts) requires connecting your site — which takes about 2 minutes and a deploy webhook or a GitHub integration.

What does "Verified" mean compared to "Advisory"?

A Verified finding means deterministic code caught a real breakage: a beacon that didn't fire, a redirect that returned the wrong HTTP status code, a flow step the browser couldn't complete. It's machine-graded — a named check produced a result, and the result is false.

An Advisory finding means the LLM flagged something that looks suspicious: a UX pattern, a consent-flow gap, a form label that might confuse users. Worth a human look — not a machine verdict. Advisory findings are never phrased as "broken." You know which tier you're reading.

Does it run automatically after every deploy?

Yes, once your site is connected. Add a deploy webhook at /webhooks/deploy — or install the Prufa GitHub Action — and every push to your main branch triggers a re-run of your money-flow checks. A Slack or email delta alert fires only when something that passed before now fails.

Will Prufa pollute my analytics or touch my real payment flows?

No to both. Analytics: Prufa runs in an isolated browser session tagged with $lib=prufa-backend — filter it out of GA4, PostHog, or Segment in your platform's filter config. Payments: checkout flows are tested up to the payment step, then stopped. No real payment instrument is ever touched. Monitoring never corrupts the data it protects.

What happens when Prufa finds something broken?

You get a Slack message or email with: the name of the broken flow, the step that failed, a screenshot of the browser at the moment of failure, and the beacon or assertion that didn't pass. The alert links to the full shareable report so you can see exactly what changed since the last passing run.

Can my AI coding agent set up monitoring without touching the dashboard?

Yes. Install prufa-mcp from the Official MCP Registry. In Claude Code, Cursor, or Cline: authenticate, call prufa.run_audit with your site URL. A 5-tool-call session sets up monitoring for any flow — no dashboard clicks required. Agents and humans get the same verified JSON output. There is no separate API mode with fewer results.

Your next deploy is tonight. Know what breaks before any customer does.

Paste your URL. A free verified audit lands in your browser in about 60 seconds.

No credit card required · No signup to run a free audit

See what Prufa finds on real sites →