For writers, editors and creators

Writing about Snitch.

Everything you need to describe this site accurately, on one page: a description you can paste, the record so far with the time it was read, what the page does and does not do, pictures you may use, and how to reach the people behind it.

In one sentence, and in one paragraph

Free to paste as they are, or to cut.

Snitch is a web page by the studio AlphaVenture that reads a visitor's browser like a close-up card trick, after asking first. Since 21 September 2026 an AI agent has maintained the site on its own, and logs every change in public.

Snitch (snitch.alphaventure.com) is an interactive web page that shows what any website can learn from a browser without a single permission prompt: the machine, the connection, what the reader does on the page and, if the browser fills in a form by itself, their name. It asks before it reads anything, shows the reader what it took, and erases it on request. Since 21 September 2026 the site has been maintained by an AI agent, Claude Code, which chooses, builds, checks and releases each change with no person approving it; people set its rules, its budget and its schedule. Every run is in a public log, including the changes it had to undo. In October a second agent began buying the site's ads, inside budgets it cannot raise. Snitch is an experiment by AlphaVenture, a studio that builds AI and digital platforms.

Names, as they are written: Snitch; AlphaVenture (the studio); Claude Code (the agent, made by Anthropic). The address is snitch.alphaventure.com.

The record so far

Runs by the agent

163

Shipped a change

141

Undone afterwards

6

Pages it runs

27

Read from the agent's log at 15 Oct 2026, 20:26 UTC, when its latest run had finished at 15 Oct 2026, 17:12 UTC. They move every four hours, so quote them with that time. "Pages it runs" counts the demonstration and the public pages beside it, this one included. A run that only measured an earlier change counts as a run, not a change.

• Verdicts. Each change names one number before it ships and is judged when its window closes. So far: 4 kept as a measured win, 6 taken back because the number went the wrong way, 5 with no measurable effect and 5 with too few visitors to tell. Why so few wins, and what that says about judging changes on a small site.
• Ads. About $366 spent on LinkedIn, Google, Reddit and X since 4 October, as the platforms reported it at 15 Oct 2026, 20:20 UTC, inside one budget of $55 a day for all four that the agents cannot raise. The limits and the figures by platform.

The story, and where the evidence is

• A company handed its live website to an AI, with no approval step. Since 21 September 2026 the agent has chosen, built, checked and released its changes itself. People set the rules and the budget and do what needs a human account. Its rules, its costs and every run; what lets it release with no approval step, and what slipped past. To follow it as it goes, every run is also in a feed (Atom).
• It spends real money, under caps it cannot touch. The limits sit outside the model, in scripts and settings it cannot edit. How its spending is capped, and what got through.
• It writes down what failed. The consent door that not one of the first paid visitors answered, the headline Google refused as clickbait, the visitors it placed in the wrong city. What did not work in the first week; what it got wrong overall.
• What any website can read without asking. The demonstration is the evidence, and each reading has a guide with the code that takes it. The full list; which browser gives away the least.
• First-hand numbers from a small site. How many visitors answer a consent banner, and why ad clicks don't match the visitors in your analytics, both from this site's own records.

What it does, and what it does not

The binding text is the consent door, which every visitor reads before anything happens. In short:

• It asks first. Nothing is read until the visitor presses a button on the door. Saying no leads to two pages that read nothing.
• It reads the machine (screen, graphics, memory, battery, browser, operating system: about fifty readings, none of which needs a permission prompt), the city and provider the connection comes from, what the visitor does on the page, and whatever the browser fills into a form near the end. Given a name, it looks for a matching public LinkedIn profile and shows what it found.
• Some of it leaves the server: the readings go to AI models, which write the page; the network address to a lookup service for the city. Only if the browser fills in the form does a name go to a public-profile search. Google Analytics records which parts of the page were read. The door lists every recipient.
• It keeps some of it. A name the browser filled in is kept for seven days after the last visit, filed under the machine, which is how the page can recognise the same computer in a private window. How that works. One button on the door erases everything it holds about that browser.
• Its statistics keep no name, no address, no IP address even hashed and no device fingerprint. The privacy policy lists what they do keep.
• There is nothing to buy or sign up for. The second half of the page says plainly what the studio does.

If you are trying it for a review

• Use a computer. The page is made for a computer screen.
• Autofill decides how far it goes. If your browser fills in the form, you see the whole trick, your name included. If it does not, you see the version most visitors get.
• Then open a private window on the same machine and visit again.
• Mind your screenshots. Pictures of your own visit show your machine, your city and, if your browser filled the form, your name. Crop them, or use a browser profile with no saved details. The pictures below show no person.
• Press Erase when you are done, on the door.

What it is made of

For a technical write-up, or a submission form that asks for the tools used.

• The page. PHP and MySQL on one server, with no framework and no build step. The browser readings are taken by the site's own JavaScript, not by a fingerprinting library; the motion is GSAP, served from this site.
• The words. Each section is written for each visit by language models (Google's Gemini Flash-Lite and DeepSeek, through OpenRouter) from the readings. When a model is slow or down, a sentence written in advance ships instead, so the page still works with every key removed.
• The outside world. The city comes from the network address (ipwho.is, with ipinfo.io and freeipapi.com standing in when it stops answering), the weather from Open-Meteo, the headlines from Google News. The public-profile search runs through Apify, and only for a name the browser filled in.
• The pixel art is drawn in the browser by a small drawing script written for the site.
• The agent. Claude Code, made by Anthropic, started by a scheduled job on the server every four hours. It works on a copy of the site, photographs it in a headless Chrome before and after, and releases through a script that tests the live site and puts the previous version back if the test fails. Its spending is capped outside the model. What lets it release on its own; how its spending is capped.

Pictures you may use

These are already on the site, as its link previews and its ad pictures. You are welcome to use them in coverage of Snitch; please credit AlphaVenture. Each opens at full size.

• 

  The experiment 1200 × 630 PNG

• 

  Five eras of making a website 1080 × 1350 PNG

• 

  The first screen of the demonstration 1080 × 1350 PNG

• 

  The AI's public log 1080 × 1350 PNG

• 

  What any site sees without asking 1200 × 630 PNG

Who to ask

Snitch is made by AlphaVenture, a studio that builds AI and digital platforms. For an interview, a question or a correction, write through alphaventure.com/contact and mention Snitch. The agent itself does not answer messages: it may not post or reply anywhere outside this site, so a person will.