What you install, and what happens next.
The relay is a shell script. It runs on your machine, probes only the connections you declared, and reports what it found — not what it concluded.
Register a fleet. You get a token.
One fleet, one scoped token. Nothing is registered yet, nothing is public, and no agent has been touched.
Takes about a minute.
Fetch your kit — it is built for your fleet alone
The relay you download is generated against your declared agents and their providers. There is no universal script, because a universal script would probe things you never told us about.
Post the token rather than putting it in a URL, so it never lands in a browser history or a proxy log.
# the kit is generated per fleet — the token is the capability curl -sS -X POST https://console.fleetregistry.ai/functions/v1/relay-script \ -H 'content-type: application/json' \ -d '{"token":"flr_…"}' \ -o fleetregistry-relay-yourfleet.shRead it before you run it
It is a plain shell script. Open it. Every probe is one visible line naming the agent, the provider, and the credential it will read from your environment.
The kit prints its own version and its own file hash on startup — so a stale copy can never quietly pretend to be current. We added that after one did.
Run it. It reads credentials; it never sends them.
Each probe asks a provider a harmless question — does this key still work — and keeps the answer. What travels back is the provider, the status code, and the provider's own error text. Your keys stay on your machine.
A provider we have no probe for is reported honestly as untested rather than assumed good.
The kit does not grade itself
This is the part worth understanding, because it is the part competitors get wrong. Earlier versions computed a verdict — verified, refuted, observed — and sent it along.
That verdict was worthless. The collector runs inside your trust domain, so a grade it awards itself is just a claim about a claim. The registry now derives state and provenance at this end, from the raw evidence. You could not fake your class if you tried, and neither could we.
Class appears — and the survey clock starts
The notation is computed. From here the evidence ages: overdue surveys redden, Conditions of Class open with deadlines, and class suspends if they go unanswered. A record that is never refreshed stops being an asset and says so.
An agent can ask to be watched harder
- Set a survey request — light, standard, deep, or continuous — before the relay runs, and the registry raises its scrutiny to match.
- Requests can only raise scrutiny, never lower it. Nothing an operator or an agent sends can loosen a survey regime; that direction is ours alone.
An agent can file a proposal without showing you its contents
- An agent may ask for something — new authority, a changed scope. The payload is hashed on your machine; only the fingerprint travels.
- The ask lands in an approvals inbox and waits for a signed human decision. The agent learns the outcome through the same channel it asked on, so the whole exchange is one chain.
Check any of this yourself.
Every claim on a public record can be recomputed on your own machine against a hash we do not hold the pen for. You do not have to take the registry's word for the registry.