Configure what your agent evaluates, launch it with a token, and let its trading fees pay for the intelligence behind it. Jev Labs is the launch and experimentation layer around Jev.
// state { "from": "0x8f…", "amount": "41,000 USDG", "age_days": 2, "prior_transfers": 0 }
An agent is a set of typed questions over whatever state you send it: pick one, rate it, yes or no. Jev answers with probabilities and confidence, in about 100 milliseconds, and never in prose.
POST state to /v1/agents/<id> and get answers your code can branch on. The questions are the ones the launcher published, signed, so anyone can check what an endpoint runs.
One transaction puts the coin on pons with a fresh fee sink as its creator fee recipient. That sink is the agent's treasury, and it lives on chain.
Creator fees accrue on the curve, anyone sweeps them into the agent's budget, and nobody can take them out. Not the launcher, not the owner. There is no withdraw.
The gateway meters every call and settles it against the budget, capped per call, with the usage files kept for audit. More trading, more intelligence, more reason to trade.
Read from the chain and from each launcher's signed config. Coverage is the floor: calls the budget buys at the cap, and the operator bills under it.
A name, and a map of typed questions. This is a risk agent for incoming transfers.
Jev{
"name": "Transfer risk",
"questions": {
"risk": {
"type": "score",
"instructions": "How risky is this transfer?",
"legend": { "0": "Routine", "1": "Unusual", "2": "Likely fraud" }
},
"hold": {
"type": "noul",
"instructions": "Should this be held for review?",
"criteria": { "true": "Hold it", "false": "Let it through" }
}
}
}
Name, symbol, an optional minimum holding to call it, an optional opening buy. One transaction on pons v2, and you sign the config so the endpoint is bound to you.
Any wallet signs once for a key. Your app sends only the state; the agent's questions ride along.
# request curl https://jevlabs.vercel.app/v1/agents/0 \ -H "Authorization: Bearer jp_0_0x…" \ -d '{"state":{"from":"0x8f…","amount":"41,000 USDG", "age_days":2,"prior_transfers":0}}' # response shape { "answers": { "risk": { "score": 1.7, "confidence": 0.81 }, "hold": { "noul": 0.88 } }, "usage": { "input_tokens": 96 } }
Measured on a fork of Robinhood Chain against the live pons factory: a launch with a 0.01 ETH opening buy, then a 1 ETH buy outside the opening window, 1% creator tax, one sweep.
No. Jev is TypeSafe's System One model. Jev Labs is a launch and experimentation layer built around it, on Robinhood Chain. Not affiliated with TypeSafe.
A decision model that does not generate prose. Send state plus typed questions, get answers with probabilities and confidence, fast enough to sit inside a request. Classify, route, score, detect, guard.
Only if the launcher gated it. An agent's door is either open, so any wallet signs for a key, or it asks for a minimum holding of the coin, checked when the key is issued and on every call.
The operator could over-report calls. That is the one trust assumption and it is bounded on chain: a cap per call, a call count and a file reference in every settlement, and the files kept for anyone to check.
The questions, yes: the launcher signs a new config and the endpoint switches to it, with the hash on the agent's page. The budget, no. Nobody can move a budget except the operator, and only to settle calls.