Announcing Wundervault 1.0 · 2026-10-27
Your agents can use your secrets. Now they can't take them.
Wundervault 1.0 puts a broker between your AI agents and your secrets. It lives on your own computer, in an account your agents can't open. Your agent asks; the broker makes the call, the payment or the signature; your agent gets the answer, never the secret.
Why we built it
Honest agents were never the problem.
AI agents now do real work: they deploy websites, look after servers and pay for services. Real work needs real secrets. Wundervault has always kept those secrets out of your agent's chat.
Before 1.0, the Wundervault program on your computer unlocked a secret when your agent needed it and used it for the agent, putting it into the command or the file the agent asked for. An agent doing its job never handled a secret. But that program shared an account with your agents, so an agent set on getting a secret could still find a way to it.
As agents take on more, and can be steered by a web page or an email they read, good behaviour isn't something to count on. 1.0 doesn't count on it.
Meet the broker
A small program with a door your agents can't open.
You install the broker once on each computer your agents use. It runs under an account of its own, with a folder only it can read. When your agent needs a secret, it asks the broker. The broker adds the secret, makes the call, and takes the secret back out of whatever comes back.
So even an agent that tries something it shouldn't finds nothing to take. The broker works the same way on Linux, Windows and macOS.
What your agents can do
Four kinds of work. No secrets handed over.
Call a web service
Any service that takes an API key or token, from Cloudflare to GitHub. Your agent says what it wants; the answer comes back without the secret.
Connect over SSH
Run a command on a server, or copy files to it. The broker signs the login, and each key works only with the servers you list.
Pay for a service
When a service charges per request, the broker pays in USDC, within a daily limit you set and to payees you trust. Agent payments
Sign something
A signature with a signing key you've added, after you've seen exactly what's being signed.
You approve what matters
The secrets that matter wait for you.
Set a secret to "Asks me first" and a call to somewhere new waits until you say yes. We email you, and you approve with your passkey: your face, fingerprint or PIN, on your phone or computer. We can't approve for you.
Signing in works the same way. One touch signs you in and, where your device supports it, opens your vault. Once you add a passkey, your password is switched off. Lost every device? A recovery code gets you back in after 3 days, and we tell you straight away, so you can stop it if it isn't you.
A new vault
Everything at a glance, and one button to stop it all.
Your secrets come first, each with what it is, which agents have it and when it was last used. Pause every agent at once, or one agent, or one secret. Give an agent a secret until a date you choose. Put back an earlier value for 30 days. Bring your secrets in from a .env file, 1Password or Bitwarden.
Help, Settings and what's new all live inside the vault, and it works on your phone.
Private by design
Your record stays with you. Your secrets stay yours.
Where each secret went, and why, is written to a signed log on your own computer, not on our servers. We keep a fingerprint of each line, so a changed log shows.
Your vault is locked in your browser before it reaches us, so we can't read your secrets. For secrets that ask you first, we keep your agent's reason only until you decide. Delete your account and everything that's yours goes at once, and from our backups within 60 days.
What's next
More your agents will do, still without the secrets.
Being built now. None of these work yet; each arrives in a release of its own.
- Coming soon
Websites, through the vault.
Your agent uses a website you've signed in to, in a browser the broker runs on your computer. It never sees your password.
- Coming soon
2FA codes.
Keep a site's two-factor setup in the vault. When a sign-in asks for the 6-digit code, the vault fills it in; your agent never sees the setup.
- Coming soon
Signing in to routers and apps.
For things with a sign-in of their own, like a home router: the vault signs in and keeps the session, and your agent sees neither.
- Coming soon
Secrets for programs on your computer.
You approve a program and where the secret goes; the vault runs it with the secret, and your agent never sees it.
- Coming soon
Admin commands (sudo).
A code that works once, in place of your password, so an agent can run an admin command you've allowed without ever knowing the password.
- Coming soon
git push
with a key the vault keeps.
Getting 1.0
New here, or moving across.
New to Wundervault
Create a vault, set up the computer your agents use, and add your first agent. It's free.
Already using Wundervault
Your account, vault, passphrase and passkeys carry over. On each computer your agents use:
- Check Node.js is 22.23.4, 24.22.1 or 26.12.1 or newer. These have the October 21, 2026 security fix, and the broker won't start without it.
- Install the new version and the broker, and pair it with your vault. Do this yourself, not through an agent.
- In your vault, open Agents and choose Repair for each agent.
- Replace the secrets your vault marks "Replace it at the service".
Thank you.
To everyone who ran the early versions and told us what broke: 1.0 is built on your notes. If something isn't right, tell us. Every change, including the ones we got wrong, is on the changelog.
About Wundervault. Wundervault keeps the secrets your AI agents need and lets them use those secrets without ever holding them. It works with Claude, Codex and any agent that supports MCP. The broker and the browser encryption are open source.