Do not trust us.
Verify it yourself.
The open agent-security stack gives defenders tools to see what an agent did. This is the layer above it: the authority that decides, enforces, and proves. Others show you a configurable simulator of a decision. Every check below runs live against the real ICA control plane, and you confirm the result yourself. No login. No account. No call to us for permission.
Verify a credential
A credential number proves a credential is real and its Ed25519 signature verifies against a published key. The organization name is a second factor, so a number alone never reveals who holds it.
Verify the ledger
Every governed action seals into an append-only chain. Ask ICA to recompute the entire audit root from the raw rows and check it against the sealed hash. You are watching the math, not our word.
Was the authority valid when it acted?
A signed receipt records a governed action. The question an auditor asks after the fact is not whether the authority is valid now, but whether it was valid at the moment the action executed. Paste a receipt hash. ICA recomputes its signature and reconstructs the authority state as of the execution timestamp, from its own timestamped records, using the same code that gates a live action.
Watch the door refuse
A vendor claiming it enforces governance is a claim. A door that refuses an anonymous caller is a demonstration. Call a governed endpoint with no credential and watch the call get denied before it runs. Then call the public one and watch it open.
Watch the kill switch
Refusing an anonymous caller is one thing. Stopping an agent that is already running is the harder one. ICA halts and quarantines a live agent, and seals the halt to the ledger before the action is stopped. Here is that state, live.
Score trust without reading a message
ICA scores the trust between two parties from communication metadata alone, never the content. The engine's safety properties are machine-checked against the real code, over every case, not sampled. Run the check. Then send your own spoof and watch the ceiling hold: with no real history, the score cannot leave the Blocked band. Trust is earned by a real, time-separated relationship, not asserted.
Every result here is live. One honest mechanic worth knowing: a freshly written seal is submitted to Bitcoin through OpenTimestamps and reads pending until a block confirms it, usually within hours. Once confirmed it is provable against Bitcoin without trusting us. Pending is what a real anchor looks like on the day it is written.
Public key, trust anchor, and the proofs: receipt public key · trust anchor · the TLA+ specifications