VetAgent

Try it

Live call against the same public endpoint your agent would use. No key, no signup.

Examples: · ·

Connect

https://vetagent.dev/mcp

Claude Code

claude mcp add --transport http vetagent https://vetagent.dev/mcp

Claude Desktop · Cursor · any MCP client

{
  "mcpServers": {
    "vetagent": { "type": "http", "url": "https://vetagent.dev/mcp" }
  }
}

Plain HTTP, no MCP

curl https://vetagent.dev/assess/0xC02aaA39b223FE8D0A0e5C4F27eAD9083C756Cc2?chain_hint=ethereum

Also listed in the official MCP registry as dev.vetagent/vetagent, on Glama, and on Smithery — same endpoint, installable from any of them.

We publish our error rates, including the ones that look bad.

Several tools in this category publish a number — Hypernative 99.8% detection, Forta >99% recall, Blockaid <0.0002% false positives (each vendor's own published figure, checked 2026-09-09). What none of them publishes is a method: no dataset, no denominator, no definition of a positive, nothing anyone can re-run. Ours is measured over 576 tokens, labelled from data sources the engine itself never reads, and the harness is in the repository.

Metric	Value	Read it as
Healthy tokens flagged high	3.1%	False positives. Too high; being worked on.
Liquid healthy tokens rated medium or high	19.5%	False blocks: what an agent refuses that it should not. One in five.
Answers returned as unknown	21.2%	Refusals when a check could not run. Honest, still too many.
Oracle-tagged centralised tokens rated high	22.3%	Most are abandoned pools holding cents, caught by the liquidity checks; most of the rest are honeypot flags. Owner powers drive none of them — the engine never scores those. USDT is rated low.
Dead tokens not rated low	86.7%	26 of 30. Read the caveat below before quoting this one.
Dead tokens rated high	10.0%	3 of 30. Our worst number, and the one a vendor would leave out.

What that row does and does not say. Recall was unmeasurable here until recently, because every public source ranks by liquidity and rugged pools fall off the list. We now recover pools from chain history — every past day is readable from the logs of the contract that created the pool — and that produced a cohort of 30 confirmed-dead tokens where sampling had produced none.

The honest reading is not flattering. Only 3 of those 30 are rated high; 11 land at medium. That is close to correct rather than a miss: 13 of the 30 dead tokens still hold $5,000 or more of liquidity, so those positions can still be sold. Dead means the project died, which is a market outcome; this tool scores whether you can get out, which is a safety property. Rating a failed investment high would be an investment judgement we refuse to make.

The number we would most like to publish — recall against deliberately adversarial contracts — is still measured on 17 tokens, of which 16 have a liquidity figure and 15 hold under a dollar (the largest holds $4,813), because the oracle that labels them raises its honeypot flag whenever its own sell simulation fails, and a sell simulation fails against an empty pool whatever the contract does. Until that cohort is larger, treat this tool as answering "can I still get out of this" and not "is this a scam". Full method and raw results →

Tools

Tool	What it answers
assess_token_risk	Should this be touched? Returns low / medium / high / unknown, a 0–100 score, and every signal that fired.
get_token_liquidity	How deep is the main pair, and can I exit at size?
find_new_hot_pools	What launched recently? Discovery only — never a safety endorsement.

Questions people actually ask

What does unknown mean?

A critical check could not be completed — usually an upstream source failed. It is not a low-risk result. evidence.data_gaps names exactly what was missing. A risk tool that guesses when it has no data is worse than one that says so.

Why do I need this if my agent can read the contract itself?

Because the failure is not in reading the contract, it is in knowing what to read. The most common loss is not an exotic exploit — it is buying a token that simulates fine but cannot be sold, or buying the wrong token with the right name. Four upstream sources disagree with each other constantly; this collapses them into one verdict.

Does it log the addresses I look up?

No. That query reveals what you are about to trade, so it is used to fetch public data and then discarded. We keep aggregate counts only — tool, verdict, coarse client name, country, and for an unknown which check could not run. Never the address, never an IP. Privacy →

What does it not catch?

Anything off-chain: team behaviour, social engineering, a rug executed through governance, a project that simply dies. It also does not yet check LP lock status or EVM holder concentration — both listed as open gaps in the scorecard. Treat low as "no fatal signal in the checks that ran", never as "safe to buy".

Is it free, and what is the catch?

Free, no key. The catch is that it is young and the false-positive rate is published above rather than hidden. We take no referral fees, no order flow, and no payment from token projects — if our revenue ever correlated with saying "low risk", the only thing this tool has would be gone.

Using it? Tell me — I genuinely cannot tell.

This server records no addresses, no identities and no token queries. That is a deliberate trade and it stands — but the cost is real: from the logs I cannot distinguish a person building something from a directory crawler, and most of the traffic is crawlers. If you have wired this into anything, one line is genuinely useful, and I will tell you before anything changes under you.

Bugs, wrong verdicts and disagreements with the benchmark are equally welcome — especially a token you think it got wrong. Those are the reports that improve it.

hello@vetagent.dev — no account needed, goes to a person.
github.com/jakegu1/vetagent/issues — if you would rather it were public, and no tracking either way.