Agent vault: secrets your AI agent uses but never sees.
An agent vault holds the API keys, SSH keys and wallets your AI agents need, and uses them on the agents' behalf. Wundervault's helper app runs on your computer: when your agent needs a secret, the helper app makes the call at runtime and hands back only the answer. Your agent is never given the secret: the AI model works with the answer, not the credential.
It works with Claude, Codex and any agent on your computer that supports MCP, the standard way agents connect to tools.
Your secrets are used at runtime, never handed over.
Each time your agent needs a secret, it asks the helper app. The helper app adds the secret, makes the call, and takes the secret out of the answer before your agent sees it.
Your agent asks for a call. It names the service, not the secret.
Your helper app adds your secret, on your computer.
The helper app sends it over HTTPS. The service answers.
Your agent gets the answer, with your secret taken out.
Your vault is encrypted in your browser, so we can't read your secrets either. Once your computers run the new version, your keys\' names and setups are locked away from us too. Secrets that matter can be set to "ask me first", so they wait for your approval before going anywhere new. Approving a request needs a passkey check from the last hour. How it's protected →
MCP tools
What your agent can do.
Call a web service
Send an HTTPS request using one of your secrets. A key can contain multiple named parts, such as a key plus an ID. Your agent chooses the destination; the helper app places each part correctly, or signs the request with it (X's OAuth 1.0a, AWS), and removes every part from replies.
vault_call
Make a new key
Ask a service for a new key, using a secret that's allowed to create keys there (Resend, Cloudflare, SendGrid, OpenAI, Twilio, OpenRouter). The helper app puts the new key straight into your vault; your agent never sees it, and no agent can use it until you save it with your passkey.
vault_new_key
Save a secret it already has
Your agent points at the file a key is in, or the command that prints it, and the helper app puts it into your vault without it passing through the chat. It waits for you, marked "Saved by an agent", until you save it with your passkey.
vault_save_secret
Set up a Cloudflare Worker's secrets
Your agent deploys a Worker that needs your database string and Stripe key. It asks once; you approve the list on your phone; the helper app puts the values straight into the Worker's encrypted settings. Nothing is pasted into the chat. The Worker's own code can read its secrets, so approve only Workers whose code you trust. See it step by step.
vault_send_secret
Connect over SSH
Run a command on a server you've approved, or copy files to it. The helper app signs the login; the key never leaves it. Admin keys can change the server, so your agent can use one only while you allow admin access for that agent.
vault_ssh_exec
Pay for a service
Pay when a service asks for payment (x402), in USDC, a dollar-backed digital currency, on the Base or Solana networks. Payments within your daily limit, to payees you trust, go ahead; other USDC payments wait for you.
Mark an SSH key as an admin key if it can change the server. It always asks you first: approve an agent's request with your passkey, or choose "Allow admin access…" on the key's page, for 5 minutes, 15 minutes or 1 hour. You can also require typing the key's name before approval.
Admin access ends when time runs out, you press End now, or the agent calls vault_done if its setting says to end access then. Its SSH connections close. On a server set up for it, the server closes them too: sign-ins use a short-lived certificate that stops working when admin access ends.
For Linux servers, guided setup on the admin key's page makes a separate everyday account, named wv-agent by default, for looking without changing. It cannot use sudo to run commands as an administrator or open the admin account's home; you can allow system logs. Setup adds health and change checks, makes keys in your browser and saves them straight into your vault; only public halves go to the server. The helper app runs a fixed setup script over SSH and shows each step as pass/fail. Nothing working is removed before a test sign-in proves the new way works, and running setup again is safe.
After each admin access ends, the helper app uses the everyday key to check for new sign-in keys, accounts, admin-group members, sudo rules, SSH settings, scheduled jobs and system services. If anything changed, you get an email saying "something changed on one of your servers", with no details. The key's page shows what changed, with Remove for a new sign-in key and "These are fine" to accept changes. A server administrator with root access can change the checker too: it catches mistakes and careless agents, not a determined one.
Setup also makes an emergency key that is never given to an agent: no copy, share or admin access. Open it with your passkey, keep a copy off the site and test it once. If the server is down, admin access cannot be turned on.
While admin access is on, the agent can read every secret on that server. Short admin access reduces that exposure; it does not remove it. An agent with its own key file to the server gets round all of this.
Without an account
One-time links
Use a one-time link only when a value must live somewhere else. For API calls with a vault secret, use the helper app instead.
Open a link someone gave you
Do not use a browser. Write the value straight to where it belongs. Never print it or repeat it in chat.
Replace LINK and PATH. LINK is the whole link, including the part after #; keep it in single quotes. If the link has no # part, send the passphrase as a second line: printf '%s\n%s\n' 'LINK' 'PASSPHRASE' (PowerShell: 'LINK','PASSPHRASE' | python ...). The opener reads both from stdin; never pass them as opener arguments.
--out PATH creates a new file with mode 600 (on Windows the file is as private as its folder, so choose one only you can open). It refuses to overwrite an existing file unless you add --force. It prints only a receipt with "burned": true, the path and the byte count. Confirm that it is burned, not what the value was.
To pipe the bare value directly into another command, use --stdout-raw instead of --out PATH. Do not let that output reach the chat or a log.
Make one for a person
Keep a generated password out of the chat. Download the same script above, then pass the value through stdin:
This prints one link, with the passphrase after the #, the same kind of link the site makes for a person. For more care, add --separate to get the link and passphrase on two lines and give them to the person two different ways. The script encrypts on your machine, using the same encryption as the browser, and creates the link with POST /api/v1/secrets. The server receives only encrypted content, never the value or passphrase.
Both commands need Python 3 and cryptography. No account or helper app is needed. A link you create this way is not a share of a vault secret.
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 helper app runs on your computer. It never sees your password.
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.
Secrets for AI agents
Agent vault vs .env files vs secret managers.
There are three common ways to give an AI agent its secrets. They differ in one place that matters: whether the secret ends up in the agent's hands.
Question
.env file or pasted into chat
A secret manager
Wundervault
Can your agent read the secret?
Yes. Anything it runs can read the file, and a secret pasted into chat is in the conversation.
Yes, once it fetches one.
No. The helper app holds it and returns only the answer.
If the agent is tricked
It can send away every secret in the file.
It can send away any secret it can fetch.
It can't read a secret, but it can direct a call, and the call carries the secret. Secrets set to "ask me first" can't go anywhere new without you.
Can the provider read your secrets?
Secrets sit in a plain file, or in the chat provider's records if pasted.
Depends on the product.
No. They're encrypted on your device.
A record of each use
None.
Of each time a secret is fetched.
Every use, with where it went and why, signed on your computer and checked against a fingerprint of it we keep.
Setup
None.
An account, plus a command-line tool or code.
An account, the MCP server, and a one-time helper app install. After that, updates are a button in your vault.
If you're an agent reading this
How to work with Wundervault
For secrets in the vault, you will never be given a secret's value. Ask the helper app to do the work. For a one-time link meant to put a value on your machine, follow the one-time link instructions.
Say why you're using a secret. The reason is kept in the helper app’s log on the owner's computer, and the owner sees it when a call needs their approval.
If a call is held for approval, wait for the owner. A call to a different destination asks again; it doesn't get around the approval.
Never ask the owner to paste a secret or a setup link into your chat.
A store for the secrets your AI agents use, built so the agents can use them without holding them. In Wundervault, a helper app on your computer makes each call with the secret and gives your agent only the answer.
Can my agent see the secret?
No. It sends requests to the helper app on your computer and gets results back.
How is it different from a secret manager?
Most secret managers hand the secret to the program that asks for it. When an AI agent can ask, it can read the secret, print it, or be tricked into sending it somewhere. Wundervault's helper app never hands the secret over: it uses it and returns the result.
Is Wundervault an MCP server?
Yes. It's an MCP server plus a helper app that runs on your computer. Your agent gets five tools: call a web service, connect over SSH, pay, sign, and list the secrets it may use.
Which agents does it work with?
Claude Code, Codex, OpenClaw, Hermes and any other agent on your computer that supports MCP.
Can Wundervault read my secrets?
No. Your vault is encrypted in your browser with keys we never receive. Only your helper app can open what your agents use, and only on your computer.
What if my agent is tricked into misusing a secret?
A tricked agent can still ask the helper app to make calls. Secrets set to “ask me first” can’t go anywhere new without your approval. Every call is logged, and the record of where your secrets went stays on your computer.
Do you see where my agents send my secrets?
Not for secrets set to run without asking. For “ask me first” secrets, you see the destination when you approve it, and so does our server.
Which secrets work?
Any HTTPS service that takes the secret as a bearer token, in a header you name, as a query parameter, as a username and password, or as part of the web address (Telegram bots, webhooks), and services that sign each request with it (AWS, SnapTrade, most exchanges). SSH keys and signing keys. And crypto wallets for x402 payments in USDC on Base and Solana, with a daily limit and payees you trust. A token exchange (OAuth) or a client certificate doesn't work yet.
What do I need?
A computer running Linux, Windows or macOS (Windows and macOS are in beta), Node.js, and an agent that supports MCP, the standard way agents connect to tools, such as Claude Code or Codex.
Is it free?
Yes. Creating a vault, adding agents and approving requests cost nothing.