Finds what gives an AI-built product away.
Ask an AI to review its own work and it says the work is good. That is not laziness. The output sits exactly where that model considers normal, so the fingerprint is invisible from the inside.
Tells is the instrument from outside. It counts what gives an AI-built product away, and it counts rather than judges. Nothing here is a model call. The same input returns the same answer twice, which is the only way a finding can carry a decision about shipping.
Human writing is uneven. Nine words, then thirty, then a fragment. Generated writing is smooth. Smoothness is the one signal that resists faking, because raising it means rewriting rather than swapping words. Tells measures that rhythm, the dash density, and the marker words.
A design is a vocabulary used consistently. A generated one is an arbitrary pile of values that happen to look fine. Tells counts the distinct spacing values, type sizes, radii and colors an interface actually uses, infers the spacing grid from the values themselves, and separates the real vocabulary from the strays.
That separation matters more than the raw count. On this domain, twelve pixels appeared 492 times and 13.6 pixels once. Read as a plain number those are two equal type sizes. They are not. One is the system and one is a slip, and telling somebody they have 23 sizes makes them give up, while telling them nine usages create four of those sizes is an instruction.
Tap targets under 44px. Buttons that overlap on a 360px screen. Elements past the viewport edge, and text cut off without an ellipsis. The layout checks need real geometry, so the server hands back a probe script your agent runs in its own browser and scores the result. No headless browser on our side, which is why this stays free.
Whether answer engines can read, quote and attribute your pages, or whether one wildcard in robots.txt has been quietly blocking them the whole time.
Computed while this page rendered, from the text above, with the module the server uses.
Everything above ships without calibration, because a count is a fact. The combined score does not, because it compares against human baselines that have not been measured yet. Until that corpus exists, Tells reports numbers and refuses to grade them.
Signal library and tests: written. Baselines: being measured. Endpoint: not answering yet.
The endpoint below is reserved and does not answer yet. It is here so the configuration is ready when it does.
{
"mcpServers": {
"tells": {
"type": "http",
"url": "https://devconnectplatform.com/api/tells"
}
}
}claude mcp add --transport http tells https://devconnectplatform.com/api/tells