Check the token, then check the transaction

A rug-check tells your bot whether a token looks dangerous. It says nothing about the transaction your bot is about to sign. That is a second, separate risk, and it is the one that empties wallets.

Two different questions

Before a bot buys a token, there are two things that can go wrong:

• The token is bad. Someone can still mint more, freeze your account, or one wallet holds enough to dump on you. A rug-check answers this from on-chain data.
• The transaction is bad. The bytes your bot signs do something else than "buy this token": they send funds to a different address, or include instructions you did not ask for. This happens when a bot builds its transaction from data it got from someone else, such as a third-party swap API, a copied route or a message from another agent.

A clean token in a malicious transaction still loses your money. A clean transaction for a rigged token does too. So the safest moment to ask is right before signing, about both.

One call: safe_buy

safe_buy runs both checks and gives one answer. You send the mint address and the unsigned transaction. We run our rug-check on the token, and pass the transaction to Presign Guard by Fizzl, which reads what it would actually do before anything is signed.

// POST https://api.agentfuel.eu/v1/safe-buy · 0.01 USDC via x402
{
  "mint": "<token mint>",
  "transaction": "<unsigned transaction, base64>",
  "intent": "buy 0.1 SOL of this token"
}

The intent field is optional but useful: it tells the transaction check what the bot means to do, so it can spot a transaction that does something else.

How the verdict is built

Each check gives a colour, and the worst one wins.

Presign Guard answers green, orange or red on the transaction itself. If one of the two checks cannot run, for example because a service times out, the answer is never better than orange. We would rather make your bot ask than let it buy on half the information.

The response also lists the reasons, so a human or your agent can see why: for example "Token: The largest holder owns 21.4% of the supply."

In your bot

The check fits in the gap between building the transaction and signing it:

Copyconst tx = await buildBuyTransaction(mint, amount);   // your existing code, unsigned
const check = await safeBuy({ mint, transaction: toBase64(tx), intent: `buy ${amount} SOL of ${mint}` });

if (check.verdict === "red") return skip(check.reasons);
if (check.verdict === "orange") return askHuman(check.reasons);
await signAndSend(tx);

Here safeBuy is a POST to /v1/safe-buy with the x402 retry shown in the docs. A malformed request gets a 400 and is not charged. For AI agents, the same check is the safe_buy tool on our MCP server, inside the free trial of 50 calls a day:

Copyclaude mcp add --transport http agentfuel https://api.agentfuel.eu/mcp

What it does not do

Green means neither check found a red flag. It does not mean the token will hold its price, or that the people behind it are honest. Both checks are automated and read what is on-chain and in the transaction right now. Use the verdict as a brake, not as a reason to buy.