Xevrion puts the right person in front of every risky agent action, lets the safe ones through, and keeps a signed record anyone can check without trusting us.
Request #4822
refund.issue for order 88231
Your rules decide who answers: nobody, one person or two. Safe actions pass straight through.
Learn moreEvery answer is signed and chained. Show an auditor one decision. They check it offline.
Learn moreSend traces from any agent. Find the actions nobody checked, and turn each into a rule.
Learn moreTest rules and prompts on real history. Block a release that makes your agent worse.
Learn moreSomebody has to answer for it. Xevrion sits between the agent and the action, so the right person says yes, the safe things move fast, and you keep proof of every decision.
refund.issue for order 88231, £540
Over £500: two reviewers must agree
Policy "Refunds" · rule 2 of 5Signed, and linked to #4821
Key 7f2c · checked offlineOne request, from the agent to the sealed record.Example data
Your agent asks first. Your rules pick who answers: nobody, one person, or two. The answer goes straight back.
Learn more →3f9a…b210a1c7…09de5be2…7c44Each decision is signed and chained to the one before it. Anyone can check one on its own, offline, with public keys.
Learn more →Decision log
Last hourResult: anySearch actions| Time | Agent | Action | Amount | Answered by | Wait | Result |
|---|---|---|---|---|---|---|
| 14:09:41 | refund-agent | refund.issue | £540.00 | PS | 6m 12s | Approved |
| 14:08:12 | support-agent | email.send | – | M | 1m 04s | Approved |
| 14:06:55 | billing-agent | db.update | – | Rule | 12 ms | Passed |
| 14:05:30 | payout-agent | payout.create | £12,000.00 | Rule | 9 ms | Refused |
| 14:03:02 | support-agent | ticket.close | – | Rule | 7 ms | Passed |
| 14:01:17 | refund-agent | refund.issue | £310.00 | P | 2m 41s | Approved |
| 13:58:44 | refund-agent | refund.issue | £880.00 | PS | 9m 30s | Approved |
| 13:55:09 | sales-agent | discount.apply | 15% | M | 4m 18s | Approved |
| 13:52:31 | billing-agent | invoice.void | £1,240.00 | P | waiting | Waiting |
Decisions per hour
Answered by peopleLet through by ruleChain
Intacta41f…09d213:41c7e0…51ab13:5209bd…e3f714:015be2…7c4414:09Each record holds the fingerprint of the one before it.
Check it yourself
Offline$ xevrion verify rec_4822 --key public.pem
✓ signature matches key 7f2c
✓ record is in the log (position 4,822)
✓ 4,821 links before it add up
verified in 38 ms, no network usedSigning keys
Public7f2cActive since 1 AugSigning3a91Retired 31 JulVerify onlyOld records stay checkable after a key changes.
This hour
Change one link and every link after it stops adding up.Example data
Replay a new rule on your own history before it goes live. See what it costs in reviews and hours. Unsafe rules are refused.
Learn more →Change a rule
12 rule "refunds"13- amount > 500 → 2 reviewers13+ amount > 250 → 2 reviewers14 otherwise → let throughReplay result
1,240 requests| Now | After | |
|---|---|---|
| Reviewed first | 84% | 92% |
| Reviews a week | 310 | 434 |
| Reviewer hours | 38 h | 59 h |
| Unsafe results | 0 | 0 |
Five versions of one rule, replayed on the same 30 days.Example data
Send traces from any agent. See each step with its time and cost, and find the actions that ran with no review.
Learn more →refund-agent · this week
Reviewed firstLet throughRefusedAlerts
2 openReviewed first, by agent
Last 48 hNoneAll
One run, step by step
run 8f31Time and cost on every step. The last one needed a person.
Rules in use
One week of one agent: how much was reviewed first, what was refused, what needs a look.Example data
Three more tools, each built on the same record.
Most teams guard agents with chat messages and trust. It works until someone asks who approved what.
See how it works →Wrap the one risky tool. It waits for a person, then runs with the approved arguments, even if a reviewer edited them.
No dependencies. Retries are safe. A crashed agent gets the same request back, not a duplicate.
Read the quickstart →// Ask a person before the refund runs const issueRefund = guard(reviews, refund, { action: "refund.issue", risk: (args) => (args.amount > 250 ? "high" : "medium"), idempotencyKey: (args, callId) => callId, }); await issueRefund({ orderId: "88231", amount: 540 }); // Waits for Maya and Jonas. Runs only if both approve.
Example code and data. The SDK and adapters are real.TypeScript
The record is built as you go, so the evidence already exists on the day someone needs it.
Xevrion helps you show control. It is not legal advice or a compliance certificate.
Read about security →The request, every decision, the ledger, the policy, callbacks, the receipt and public keys in one file.
A signed report for a period: control statements, policies, outcomes, reviewers and checked trails.
Plain signals with the numbers behind them, signed, and sealed to your own key if you want.
Push every ledger entry to your SIEM, signed, as NDJSON or CEF. Content-free: payloads are hashes.
Set how long content is kept. Erase one request on demand. Legal hold stops it. The proof stays valid.
Single sign-on, SCIM, custom roles, and a second owner for protected changes.
Pin data to the UK, EU or US. Encrypt with your own key on Enterprise.
Spend and request limits that step up as an agent earns trust from human outcomes.
Questions about your own review? Email hello@xevrion.co.ukWritten enquiries only
The same record, seen the way each team needs it.
Every plan includes signed proof and the reviewer inbox. Annual billing costs ten months, so you get two free.
Compare every feature →Try it with one reviewer.
1 reviewer
Get started freeA first agent in production.
3 reviewers. Or £990 a year.
Get startedTeams running several agents.
10 reviewers. Or £2,490 a year.
Get startedRegulated and larger teams.
Seats and volume agreed with you
Write to usBilling is not switched on yet, so no plan charges today. Prices may change before it is.Pricing
Still stuck? Email hello@xevrion.co.uk. A person replies.
A review and proof layer for AI agents. Your agent asks before risky actions, a person decides, and every decision is kept in a record anyone can check.
You add a few lines before the risky action, using the SDK or the API. The rest of your agent stays as it is.
Only for actions your rules send to a person. Everything your policy approves automatically goes straight through.
No. Your organisation's data is separate, and you can lock it with your own key on Enterprise.
In the region you pick: UK, EU or US.
No. It helps you collect evidence and maps it to common frameworks. It is not legal advice or certification.
Yes. Start free, with one reviewer, and move up when you need more.
Write to us. We answer in writing, by email.
Start free. Guard your first agent in a few minutes, and keep a record anyone can check.
Signed decision
Checked on your own machine