Give an agent a token. Wake up to a report where every number has a receipt. No buy or sell calls, ever.
Try it out with a sample
Pick a creature, write what it should find out, and set its limits: steps, cost per run, a daily budget.
It reads the chain (holders, pools, the creator's wallet) and opens pages through a separate reader, never raw.
Code rules and a separate judge check the report against its own data before it reaches you.
A morning digest, not an endless feed. Mark reports reviewed, dismiss the rest.
A report may only use numbers a tool returned in that run. Anything invented is rejected by code, not by asking the model nicely.
Each number in a report points at the exact tool call it came from. The tool results are hashed, so the report can't be quietly changed later.
Hover or tap a number to see the step it came from.
Every step is public, with a short reason the agent gave. Nothing happens off the record.
A cheap reader model turns each page into notes. The thinking model never sees raw page text, so a page can't give it orders.
The second look grades the report, not the token. One question: do the claims match the data in the receipt?
Research people chose to publish, with its receipts and its second look. Including the reports that didn't hold up.
Pick a metric and a threshold. Watches run on every pass without spending a cent on a model, and an agent can wake up when one fires.
This is the same check the app runs on every token an agent watches. Try it on three sample tokens.
2 of 3 met · user-defined
Pool Fox woke upAn agent can sleep until the numbers move. When a watch fires, it runs and tells you why.
model cost $0.000
PASS · EVERY RUNWatches are plain comparisons in code. They never call a model, so they cost nothing to run.
Pick from a fixed set. No scripts and no prompts, so there's nothing to break or abuse.
Agents run overnight on a schedule. In the morning you get one short list: what they found, what held up, what didn't.
These aren't instructions we hope the model follows. They're checks in code that run before anything reaches you.
✕ tells the reader to trade now
SAFETY · TEXTBuy, sell and hold instructions, price predictions and "scam" labels are rejected.
Off by default. If you let an agent handle funds, it works inside limits set on-chain: how much, and where to. You can revoke it any time.
run cost cap $0.05
■ stopped at cap · try not used
Steps, cost per run and a daily budget. A run that hits its cap stops, and doesn't use up a try.
Free while we're in early access. We pay for the research.