A terminal replaying a Leakdown run: it reads the site, maps it, builds ten prospects, sends them through signup, filters what more than one cites, and writes the report.
You are losing customersin the new web era.
We send a spawn of agents through your site in a real browser to check your
flows, catch where visitors drop, and check how ready you are for agents.
git clone https://github.com/0xSarnavo/leakdown-cliCatch them before you lose them.
Leakdown is in alpha. A simulated prospect stalling is a signal, not a measurement of your traffic. Every finding carries its count, and under three sessions it says too few to call. Check it on your own page before you ship a fix.
Find the leak before it costs you a customer.
Leakdown sends prospects and AI agents through your site in a real browser, and tells you exactly where they got stuck.
- 01WHERE THEY WALK OUT
Watch
Analytics show where people drop. Leakdown shows why. Ten prospects drive a real browser, think out loud, and say the moment they give up.
- 02AGENTS AT THE DOOR
Unblock
AI agents now sign up for things. A CAPTCHA, an SSO-only gate or an unlabeled button stops them, and you never hear about it. We hit those walls first.
- 03EVERY DEPLOY
Guard
A deploy can break signup and nobody writes in. Run one goal in CI and get a pass or fail before a customer finds it.
- 04WHAT TO FIX FIRST
Fix
One number, the three pages people left from, and an expert fix for each, with steps to check it yourself.
Request a run
FOUR CHECKS · ONE RUNREQUEST A RUN >>>
Frequently asked questions
It is a free early-access demo: we run a full run (or the special run you asked for) on our side and email you the report. We collect your site URL, your email, your run choice and brief, and whether you allow public sharing. We keep the run's videos, logs and screenshots of your site with the report. Nothing is sold, and everything is deleted when you ask. This site has its own analytics, described in the privacy policy; the site we test for you is never tracked. Findings go public only if you opt in, and only once they are fixed.
Only run it on sites you own or have written permission to test. It creates real accounts and triggers the real emails and webhooks a signup triggers, so use a staging copy when you can.
A guard refuses three things at the action itself: paying, confirming a booking, and signing in with Google, GitHub or SSO. Reaching that wall is the finding. The matching is best effort, so don’t point it at a live checkout and assume it can’t buy. Four more — deleting data, inviting teammates, publishing anything, and opening support chat or contacting third parties — are a rule the prospects are told to follow, not machinery: a “request a demo” form or a trial that needs no card can still be submitted.
Everything lands under runs/<site>/ on the machine that ran it — nothing goes to us. Your AI CLI does see each page your prospects visit, and any verification emails they read, under that provider’s terms. Delete a site with rm -rf runs/<site>. Requested runs keep your URL and email privately, only to send the report.
The CLI is MIT and runs on the AI CLI subscription you already have (Claude Code, opencode or Codex), with no API keys — check that provider’s terms and rate limits for automated use, and expect a long sweep to hit a personal subscription’s limit. Requested runs are free during early access, one per day, and run at our discretion.
The full --ladder run takes about an hour. A single --goal check is a handful of steps, made for CI: exit 0 pass, 1 fail, 2 could not run.
Read it as risk, not traffic. A simulated prospect stalling is a signal that real visitors could, never a measurement of them. Every finding carries its count, and under three sessions it says too few to call.
Run it yourself,
or have us run it.
or run it yourself: git clone leakdown-cli ↗