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
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
"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."
"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."
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.
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.
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.
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.
Runs meter one LLM call each; confirmed-flow replays are never metered. Full pricing & billing details →
Answers in plain English. No marketing speak.
/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.
$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.
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.
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