onchain_agent_preflight

USE WHEN you are about to install a package, clone a repo, connect to an MCP endpoint, pay an agent, or trade a token, and you want to know what is on record about it FIRST. Pass exactly one of repo, package, endpoint, agent or token.

Returns a verdict — go | caution | no | unknown — with one evidence line per check, each naming the field it was read from and when that field was written. The rules are written down in lib/preflight.ts and cited by id in rule.

RULE ENFORCED: a verdict names WHAT WAS CHECKED AND WHEN. It is never a security review, a quality judgment or a statement about returns, and unknown means Sato Hub holds no record — not that something is wrong. An unlisted endpoint gets ONE live handshake (initialize + tools/list, 8 s cap) and can never come back go: a handshake is not a record. For agent=<chain>:<id> we confirm the ERC-8004 registration exists, fetch its registration file, and report the services it DECLARES; only a declared MCP service is probed.

TOKEN LANE (token + chain, EVM only): keyless chain reads — bytecode presence and size, the ERC-20 views, the Clanker v4 factory's OWN deployment record (tokenDeploymentInfo, not a bytecode heuristic), and the Uniswap v3 factory across the four standard fee tiers against wrapped native. Every field is nullable and a null carries the reason it is null. PERMANENTLY NULL, and said so in the evidence: holder concentration (no keyless public source — explorers are not scraped) and Uniswap v4 / non-Uniswap liquidity (a v4 poolId cannot be reconstructed without the PoolKey). The deployer address needs an optional explorer key. A pool existing is not depth; a locker holds a position on the terms its own code enforces. Nothing in this lane says safe, audited, rug or scam — those are not readings.

SKILL LANE (skill): a skill is a DOCUMENT an agent follows, which is exactly why it is worth checking first — the ClawSwarm skills needed no malware, only text telling the agent to generate a wallet and post the private key. Evidence is the static disclosure the weekly sweep already produced: hosts the text names, whether it generates or handles keys, whether it asks for a credential, whether it pipes a remote script into a shell, what tools it grants itself — each finding WITH the lines that produced it — plus installs, when it was last seen in its registry, and whether a host it names belongs to a listed project. Nothing is fetched from a registry and no skill is executed. A DISCLOSURE DESCRIBES: it never says safe, it never says malicious, an empty flag list is "nothing matched" rather than a pass, and the registry's own scan result is attributed to that registry by name.

Returns (json): { verdict, rule, target: { kind, value, slug, name, sato_url, verify_url }, evidence: [{ check, result, source_field, checked_at }], checked_at, caveat, rules, token?, skill? }. token and skill are the raw reports for those lanes. Read-only.

Example: { repo: "coinbase/agentkit" } · { endpoint: "https://mcp.example.com/v1" } · { agent: "base:42" } · { token: "0x1bc0c42215582d5A085795f4baDbaC3ff36d1Bcb", chain: "Base" } · { skill: "clawhub/solana-wallet" }