Security

Last updated 21 November 2026

Xpenvo Agents sits between your agents and your AI provider, so we built it to hold as little as possible and to fail safe. We never store your prompts, your answers or your AI provider keys. If anything needed to enforce a budget is missing or broken, the request is blocked, never let through unchecked.

Your requests and keys

Prompts and answers are never stored

They pass through like water through a pipe. We read only the token counts, to work out the cost. Request records hold metadata only, and are deleted after 90 days.

Your provider keys are never kept

Your AI provider key travels with each request, is forwarded, and is forgotten. It is never saved and never logged.

Keys only go to the provider

The proxy talks only to api.openai.com and api.anthropic.com, and never follows a redirect, so your key cannot be bounced somewhere else.

Agent keys are fingerprinted and signed

We keep only a one-way fingerprint of each agent key. Every key carries a signature only our server can make, so made-up keys are refused at the door.

Fail closed

An unknown model, a missing setting or an unreachable budget engine means the request is blocked with a clear error, never sent on unchecked. If our budget check cannot answer, your agent gets a 503 error and nothing reaches your provider, so an outage on our side can never spend your money.

Cut-off replies are counted in full

If a reply stops early, for example when your agent hangs up mid-stream, and the provider did not report its usage, we charge the full amount we set aside before the request. Budgets are never under-counted and nothing stays held.

Each customer is walled off

Every organisation has its own budget engine with its own private database. A key from one organisation cannot reach another’s.

Your account

No passwords

You sign in with a one-time code sent to your email, or with Google or GitHub. There is no password to steal, guess or leak.

Google sign-in comes straight back to us

Google sends you back to agents.xpenvo.com only. Three one-time checks tie each Google sign-in to the browser that started it, so a stolen or forwarded sign-in link is useless. We ask Google for your email and nothing else.

Two-factor sign-in

Add an authenticator app so a stolen email account is not enough. Organisation owners can require it for every member.

Sessions scripts cannot steal

Sign-in passes live in secure cookies that page code cannot read, sent only over HTTPS to our exact address. Sign out on one device or everywhere at once.

Roles checked on the server

Owner, admin and viewer rights are checked by our server on every request, not just hidden in the page. Removing someone takes effect on their next click.

New keys are shown once

A new agent key appears once, for you to copy. It is never saved in your browser, and we keep only a fingerprint of it, so it cannot be read back later by anyone.

Safe downloads

The activity file you download for a spreadsheet is plain text. Anything that looks like a spreadsheet formula is marked so it is shown, never run.

Every change is in the audit log

Who changed what, and when, for every budget, key, member and setting. Kept for one year.

Nobody can probe who our customers are

The sign-in page answers the same way for every email address.

Against automated and AI-driven attacks

Attackers now use AI to run huge numbers of tries from thousands of addresses at once, and to write convincing fake emails. We designed for that.

Guessing limits that hold worldwide

A sign-in code or authenticator code gets 5 tries in 15 minutes, counted once for the whole world. Spreading an attack over thousands of addresses in many countries gives no extra tries. Each email address gets at most 10 codes an hour.

Every try counted first

Tries are counted before the code is checked, so a burst of thousands of tries sent at the same moment gets no further than 5.

Nothing to guess or crack

There are no passwords. Codes expire in 10 minutes and work once. Agent keys are long random secrets with our signature, and our admin tools lock out any address that keeps getting the admin token wrong.

Fake emails get nowhere

Your sign-in code is only ever typed on agents.xpenvo.com. We will never ask you for a code, an authenticator code or an agent key by email, chat or phone. If anyone does, it is not us.

One address for sessions

Sign-in and your account only work on agents.xpenvo.com. A copied session cannot be used anywhere else, and other websites cannot act for you.

Few moving parts

The proxy uses no outside code libraries at all, and the website only three. Every week, we are automatically told about security fixes for them.

Our systems

Encrypted everywhere

All traffic uses HTTPS, and browsers are told never to use plain HTTP for our site. Data is encrypted at rest by our providers.

Rate limits

Sign-in attempts, account actions and each agent key have per-minute limits, so a leaked or buggy agent cannot flood the service.

Locked-down website

Our pages run only our own code, cannot be shown inside other websites, and refuse changes sent from other websites.

Minimal logs

Automatic per-request logs are switched off. Our own error lines hold a request id and an error type, never content, keys or email addresses.

Secrets stay secret

Server secrets live only in our hosting provider’s encrypted settings, never in code, on GitHub or on the website.

Safe database code

Every database command is written in advance, and input is only ever passed in as a value, so it cannot change the command.

Speed

The checks run on Cloudflare’s network, close to your agents, before the request leaves for your provider. Every reply carries a standard Server-Timing header, for example xpenvo;dur=4, showing how many milliseconds our checks added, so you can measure it yourself instead of taking our word for it.

Who helps us run the service

ProviderWhat it doesWhat it can see
CloudflareHosts the website, runs the proxy and stores budgets, key fingerprints, request records and audit logsRequest metadata. Prompts and answers pass through its network but are never stored
SupabaseSign-in, two-factor sign-in and the waitlistYour email and sign-in records
ResendSends our emailsYour email and the email text
Google and GitHubSign-in, only if you choose themThat you signed in to Xpenvo
Dodo PaymentsPayments, tax and invoices as merchant of record, only once paid plans start after the betaYour organisation id, the plan, and what you enter on its payment page. We never see card or bank details

We tell customers by email at least 30 days before adding a new provider that handles personal data. Our Data processing terms cover how we act for your company.

Your data, your control

You can download everything we hold about you, or delete your account, from your account page at any time. Organisation owners and admins can export the whole organisation. Details are in our Privacy policy.

Where it runs

The website and proxy run on Cloudflare’s global network, close to your agents wherever they are. Sign-in is handled by Supabase and emails are sent by Resend. The full list, and what each provider does, is in the Privacy policy.

Report a security issue

If you think you have found a vulnerability, please email [email protected] with the details and how to reproduce it. We will reply within 3 working days and keep you updated until it is fixed. Please do not access other people’s data, disrupt the service or share the issue publicly before it is fixed. We will not take action against anyone who reports in good faith under these rules.