AI trading agents on Base are increasingly autonomous — they see a
token, decide, and execute, often with no human in the loop. But
"decide" usually means trusting whatever on-chain data the agent
can scrape in the few hundred milliseconds it has before the trade
window closes. That's a thin foundation for judgment: honeypot
contracts, unlocked liquidity, rug-pull patterns — none of that is
visible from a raw balance check.
Most existing risk tools are built for humans: a dashboard, a
search box, a subscription. None of that works for an agent that
needs a machine-readable answer in one HTTP call, with no signup
flow, no API key provisioning, no human clicking "I agree" anywhere
in the loop.
AgentRisk is built for that gap specifically — an m2m risk oracle,
not a dashboard.
How it works
Pay-per-call via x402. No subscription, no key. An agent sends
a request, gets a 402 Payment Required response, pays 0.15 USDC on
Base, and the same request goes through. No account creation
anywhere in the flow — which matters when the caller is a bot, not
a person who can fill out a signup form.
Cross-referenced signal, not a single source. Every scan
combines GoPlus token-security data with independent on-chain
checks — LP lock status, holder concentration, sell simulation
against the live pool. A token can look clean on one source and
show a blocked sell path on the other; AgentRisk surfaces both
rather than picking one.
Signed, cacheable verdicts. Every response can carry a signed
attestation: risk score, scan id, rulepack hash, policy version,
input digest, chain, timestamp. This came directly from a
conversation with another agent builder (@AxiomBot / Protean
Labs) who pointed out that agents calling this at scale need to
cache a verdict and trust it without re-querying — the receipt is
what makes that safe. An agent can verify the signature offline,
check the timestamp isn't stale, and skip the network call entirely
on a cache hit.
Machine-readable failure states. A signature check doesn't just
return "invalid" — it distinguishes expired, revoked,
replayed, and unsigned, so a calling agent can decide
programmatically whether to retry, re-scan, or hard-fail.
Integration surface
-
REST endpoint —
POST /scanwith a contract address, x402 payment header, get back structured JSON. - MCP server — for agents built on Claude/MCP-compatible stacks, risk scoring is exposed as a tool call.
- A2A Agent Card — so other agents can discover the service and its capabilities without hardcoded integration.
Where it stands
The service has been live on Base Mainnet for a few months. It's
listed in a couple of community agent-tooling directories
(awesome-a2a among them) and has processed real scans end to end,
including real x402 payments settling on-chain.
It's early — real usage is still small, and that's an honest
statement, not a hedge. If you're building an agent that trades
autonomously on Base and want pre-trade risk data without wiring up
your own token-security pipeline, the endpoint is open:
agentrisk.dev
Feedback — especially from anyone building m2m agent infrastructure
— is genuinely welcome. The attestation spec above exists because
someone in a Twitter thread asked a sharp question about staleness
and caching; more of that kind of pressure-testing is exactly what
this needs right now.
Top comments (0)