How it works
Authority, action, evidence, verification.
Covera turns a consequential AI action into something an outside party can check. You do not have to change your AI stack, and nobody needs an account to read the result.
1. An AI receives authority
A person, a company or another system says what the agent is acting on behalf of and what it may do. This is recorded and dated before anything runs, which is what makes it evidence rather than an explanation.
2. The AI takes a consequential action
Something changes outside your systems. Money moves, data is edited, a message goes out, a tool runs.
3. Evidence is collected
Identity, authority context, what was done, the amount, approvals, the result and the time are joined into one record.
4. Covera verifies the evidence
Integrity first, then the authority the action was working under, then whether anything is missing. Each check is reported in plain words, pass or fail.
5. The result can be shared
Anyone with the link can inspect the evidence without having to take your word for what happened.
Write down the authority first
Name the agent, list the actions it may take, set the most it may spend at once, and say when a person has to approve. Covera stores that with a date, so it cannot be rewritten after something goes wrong. Without it, nobody outside your company can establish whether the agent went too far.
Send the action
One call, at the moment it happens. Nothing else in your code changes.
curl -X POST https://covera.app/api/public/v1/actions \
-H "Authorization: Bearer cvr_live_your_key" \
-H "Content-Type: application/json" \
-d '{
"agent": "refund-agent-14",
"action": "refund.create",
"principal": "Example Company",
"amount": 240,
"currency": "USD",
"human_approved": false,
"result": "completed"
}'Get something you can send to anyone
{
"receipt_id": "CVR-8HF92K",
"verify_url": "https://covera.app/r/8HF92K",
"verdict": "verified",
"evidence_hash": "9f2c...",
"findings": [
{ "label": "Agent identified", "detail": "Refund Agent #14 is registered and active." },
{ "label": "Authority", "detail": "USD 240 is inside the recorded limit of USD 500." }
]
}The page shows the agent, who it acted for, what it did, the authority it was working under, whether a person approved, and whether the record has changed since. The reader's own browser recomputes the fingerprint while they look at it.
Fields
- agent
- The short id you gave the agent. Required.
- action
- What it did, like refund.create or email.send. Required.
- principal
- Who it acted for, usually your customer.
- amount and currency
- Money involved, if any. Checked against the authority.
- human_approved and approver
- Whether a person signed off, and who.
- result
- completed, failed, or your own word.
- details
- Anything else you want on the record. No customer data is required.
- occurred_at
- When it happened. Defaults to now.
What Covera does not claim
Covera does not decide who is legally liable and does not say an AI action was the right call. It establishes and checks the evidence that people and systems need to make those decisions themselves.
What Covera never sees
No prompts. No model answers. No customer records. No access to your systems and nothing read from your accounts. You send the shape of the action, not its contents, which is what makes this safe to switch on across a whole company.