How it works

Every launch says it does something: a launchpad, an AI agent, a scanner, an app. NOCAP asks one question: does it actually do that? Four AI agents go and check, in parallel, and you watch them work. Then a judge grades it, with the receipts.

Three ways to ask

Here: paste a CA, the project's site, or an X post. A check takes about a minute; the same target is cached for an hour.
On X (at launch): reply to the bot under any project post. It answers with the verdict, one line and a link that unfurls as the card. It only answers when tagged and stops for anyone who writes "stop".
Caught today: every new launch with its own site that crosses $10k goes through the same four agents automatically. You can watch those runs live too.

The four agents

Site agent Claude Haiku 5.5

Uses the product in a real headless Chrome, the way a buyer would.

First checks the site is the project's at all: redirects, a parked page, a platform, an old domain that never mentions the coin. A borrowed site is a receipt; nothing on it is tested as the product.
Then: closes pop-ups, screenshots, watches "live" numbers, follows the main button, tries the core action (a CA into a scan box, Play on a game), reads the site's own API.
Then the model takes the browser: up to 10 clicks, inputs and reads to test what the scripted pass couldn't see. Every screenshot streams to the console.

Chain agent Claude Haiku 5.5

What the chain says about the coin behind the project.

The pool, the mint, the launch block (who else bought in the same block), the top 20 holders.
The dev wallet: its earlier coins in our radar, 40 txs since launch.
The CA printed on the site against the mint; claimed burns against the supply. The model may read up to 2 wallets the project names (treasury, fees, burns).

Social + code agent Claude Sonnet 5.5

Who is behind the claim, as far as public accounts show it.

The X account: age, followers, posts, the pinned and last 10 posts. A big old account that never mentions the coin is a borrowed link.
The linked repo: commits, age, and the code itself. Sonnet reads the files that would implement the main claim and records what it finds, with a quote we check against the file.
Docs pages and store listings the project links. Haiku does the reading when there is no code.

Judge Claude Sonnet 5.5

Reads every receipt from the three agents and writes the result.

Sorts the project's own claims into held, failed or not seen, citing receipts.
Writes one line. A validator throws out any line with a number, name or quote that isn't in the receipts.
The score itself is plain arithmetic over the receipts, itemised on every check (below).

Models never decide a receipt: every receipt comes from what a tool actually returned (a page, a transaction, a file). Each check has a hard cost cap and the whole service has a daily one.

The verdict

CAP0–20. A claim turned out false: a fake or dead app, fake burns, a wrong CA, a borrowed site or X, a missing repo, faked code, or its own words say the core isn't live.

MOSTLY CAP21–40.

HALF CAP41–60. Part works, part doesn't.

MOSTLY NO CAP61–80.

NO CAP81–100. It does what it says.

JUST A MEMENo score. A meme, lore or community coin that pitches no product and says nothing false. Shown with its chain basics (launch block, the dev's earlier coins, the CA on its site) and the live market. A meme only goes on the CAP scale when something it says turns out false.

TOO SMALLNo score. Under $10k of live mcap (or a dead pool) the AI agents aren't spent: you get the live mcap and volume and the chain basics, and "check again past $10k".

Every project gets a verdict. When there's no app to use (no site, a wallet connect, a captcha, a beta), the verdict rests on the other agents: the chain (the CA, a program it prints, burns), the X account and the code. With nothing failing, a project can't go under HALF CAP: our browser not getting in is never held against it. With nothing failing and real receipts behind it (a deployed program, a working button, a real repo, its own API answering), it starts at MOSTLY NO CAP; with its core claim seen working, at NO CAP. Bundles, top holders and the dev's earlier coins are context: they move the score a few points and never make a CAP.

Market context on every check: live mcap, 24 h and 1 h volume, liquidity and age from DexScreener (a pool under $1k liquidity is called dead, never priced), else our own read of the pump curve or pool on-chain, else "price unavailable". Never pump.fun's cached mcap or ATH.

How the score is made

Every check starts at 40 and moves with what the agents found. These are the points, read live from the code:

Seen	Points
Loading…

A product partly behind a wallet connect stays at or under 80. A site is only called down if a second network can't open it either; then it's capped at 15.

What the browser never does

No wallet, ever.: It never connects a wallet, never signs, never presses buy, swap, send, deposit, mint or claim. Popups are blocked.
No personal fields.: It never types into email, password, seed or payment fields.
Public sites only.: Links that point at private networks are refused before and after DNS.

Holder pass

Every check is free. Holding $NOCAP gets you 60 checks a minute instead of 12, and, if you add your X handle, the bot's priority lane for 24 h. You prove it by signing a plain message: no transaction, nothing moves. No payouts, no points, no promises.

Switches on when the CA is live.

What it isn't

Not a team audit.: It checks the claim, not the people. A project that does what it says can still fail or rug.
Not a call.: NO CAP means the product worked when we tried it. It doesn't mean the coin goes up.
Logins and waitlists are judged on what is public.: Anything behind a login, a wallet or a beta is judged on its repo, socials, CA and coin. Not seen working is not proven fake.
Missing data is said out loud.: If a source doesn't answer, the receipts say what wasn't measured.

FAQ

How long does a check take?: Usually 40–80 s: the three agents run at the same time, then the judge. The live median is in the bar on the front page. The same project checked again within an hour answers at once from the cache.
What if the app can't be used?: Then the verdict rests on the other receipts: the repo, the X account, the CA on the site, the dev wallet and the coin. A claim of a long history with a repo made this morning fails; a real repo with months of commits counts. Nothing is called CAP just because our browser couldn't get in.
When is something called CAP?: Only on hard evidence: the site is gone and a second network agrees, the site itself says the core isn't live, an answer is hardcoded (no request leaves the page), a linked repo or store listing doesn't exist, the site shows a different CA, or two independent things are broken.
Does it touch my wallet?: No. You never connect anything to check, and our browser never connects a wallet, signs or pays on the project's site.
Can a project game it?: A project can make its product work. That's the point. Every step each agent took is in the console on the check page, with screenshots and receipts, so a verdict can be checked by anyone.
Where does the coin stamp come from?: Our coin engine: pool and volume against the cap, the dev's create buy and sells, launch-block bundles, holder concentration and copies, from the chain and our radar. It sits next to the project verdict, not inside it.
Is a NO CAP a buy signal?: No. It means the product did what it says when we tried it. Not financial advice.