@wundervault/mcp-server

A zero-knowledge secrets vault for AI agents. Every API key you paste into an agent chat or a .env file ends up in context windows, transcripts, and provider logs. Wundervault's answer: the agent never receives the secret at all. It asks for work — "run this deploy with the key injected" — and a local daemon decrypts the secret, injects it into the subprocess environment, zeroes the buffer, and scrubs the output before the agent sees any of it.

This repo is the MCP server that exposes that workflow to any Model Context Protocol client — Claude Code, Cursor, Cline, and others.

Don't trust the claim — test it: the zero-knowledge property is independently verifiable at your own network boundary in about 5 minutes (browser DevTools or a mitmproxy canary test). Guide + our own test transcript: wundervault.com/verify.

How it works

┌──────────────┐  MCP (stdio)  ┌───────────────────┐  ciphertext only  ┌───────────────────┐
│   AI agent   │──────────────▶│  wundervault-mcp  │◀─────────────────▶│  wundervault.com  │
│ (Claude, …)  │◀──────────────│  + local daemon   │                   │ stores encrypted  │
└──────────────┘ "burned" ack  │  decrypts HERE    │                   │ blobs, no keys    │
                               └─────────┬─────────┘                   └───────────────────┘
                                         │  secret → subprocess env
                                         │  (buffer zeroed after spawn)
                                         ▼
                               ┌───────────────────┐
                               │   your command    │ stdout/stderr scrubbed
                               │ (deploy, API, …)  │ before the agent sees it
                               └───────────────────┘

Secrets are encrypted client-side (AES-256-GCM via Web Crypto) before upload. The hosted service only ever stores ciphertext — it cannot derive the key, the passphrase, or the plaintext.

Install

npm install -g @wundervault/mcp-server

Quick Start

{
  "mcpServers": {
    "wundervault": {
      "command": "wundervault-mcp",
      "env": {
        "WUNDERVAULT_AGENT_NAME": "<agent-name>"
      }
    }
  }
}

Keys are never placed in the MCP config. The server names its agent, then asks the local wundervault-agent daemon for that agent's credentials over a unix socket. onboard.py registers the agent and starts the daemon.

New account? wundervault.com has a 90-second agent onboarding flow that generates this config for you.

Supported platforms

Linux is the verified platform. macOS works for secret delivery; Windows does not.

Delivery is POSIX-only by construction: the sudo recipe pipes through /bin/sh, and the git / ssh-passphrase recipes need mkfifo and setsid. On Windows those mechanisms return a clear "not supported" error rather than failing somewhere deep inside.

The single-instance lock has two implementations: a kernel-held abstract socket on Linux, and loopback ports elsewhere. Only the Linux one is covered by tests — see HANDOFF-lock-on-non-linux.md for what is unverified on macOS and why. CI runs both platforms; the lock suite runs on Linux.

Security Model

• Zero-knowledge: The encryption key lives only in the MCP server process. The Wundervault server never sees it.
• Burn-after-reading: Plaintext secrets are never returned to the calling agent. After decryption, the agent receives only "Secret retrieved and burned.".
• Exec scrubbing: Command stdout/stderr are scrubbed of the plaintext before being returned; shell-escape patterns ($(), backticks, sh -c, eval) and file redirects of secrets are rejected before decryption.
• Directive integrity: Server-side directive signatures (PBKDF2-HMAC-SHA256, 600k iterations) are verified before any secret is released.
• Timing-safe: HMAC comparison uses crypto.timingSafeEqual.
• Tiered access: Per-entry access tiers are enforced server-side; high-tier secrets require human approval before an agent can use them.

Honest limitations

• The platform is open-core: this MCP server and the browser crypto are AGPL-3.0 so you can audit everything that touches your secrets, but the hosted service itself is not open source.
• A local daemon must run next to the agent; fully air-gapped setups don't fit.
• By design the agent can never read a secret's value — if your workflow needs the model to reason about the secret itself, this is the wrong shape.

Tools

vault_entries_list

List all vault entries available to this agent. Returns entry IDs and secret names — no values.

Input: {}
Output: "Vault entries (N):\n  [entry_id]  secret_name  (tier: read)"

vault_entry_get

Retrieve and decrypt a vault secret. Optionally execute a command with it.

Input:
  entry_id: string          # from vault_entries_list
  purpose: string           # audit log reason
  exec?: string             # optional shell command

Output: "Secret retrieved and burned." (plaintext NEVER returned)

Secure exec pattern (sudo example):

sudo -S systemctl restart nginx <<< "$WUNDERVault_SECRET"

Do NOT use echo $WUNDERVault_SECRET | sudo -S — that exposes the secret in process logs.

vault_exec

Execute a shell command with a vault secret injected as an env var — locally or on a remote host over SSH. The secret is injected into the subprocess and the buffer is zeroed immediately after spawn; escape patterns are rejected before decryption.

Input:
  purpose: string           # audit log reason
  command: string           # full shell command (no escape patterns)
  entry_id?: string         # secret to inject (omit for SSH-key-only remote exec)
  working_dir?: string
  inject_as?: { env_key, pre_command?, post_command? }   # override entry's exec_config
  remote_host?: { host, user, ssh_key_entry_id? | ssh_key? }

With remote_host.ssh_key_entry_id, the SSH key is fetched from the vault and used without ever being written to disk.

vault_entry_inject_env

Write a vault secret directly into a config file (~/.npmrc, ~/.netrc, ~/.docker/config.json, or a project .env) without the plaintext passing through the agent.

Input:
  entry_id: string
  purpose: string
  file_path: string         # allowed config file paths only
  env_key: string           # variable name to set

vault_rsync

Sync a local directory to a remote host using rsync over SSH, with the SSH key fetched from the vault (temp keyfile deleted immediately after transfer).

vault_entry_forget

Discard a local reference. No-op on the server.

Input: { entry_id: string }
Output: "Reference [id] discarded from local context."

Credentials

The MCP server holds no keys of its own and takes none on the command line. On the first tool call it resolves them like this:

1. WUNDERVAULT_AGENT_NAME (required) names which registered agent this process is.
2. The agent token is read from WUNDERVAULT_AGENT_TOKEN, or from ~/.wundervault/agents/<name>.token.
3. That token is presented to the local daemon over ~/.wundervault/agents/<name>.sock, which returns the API key, the encryption key, and the vault URL.

If the daemon is not running, tool calls fail with instructions rather than falling back to a weaker source. Run onboard.py to register an agent and start it.

CLI Options

wundervault-mcp [options]

  --url <url>   API base URL override (default: supplied by the daemon)
  --help        Show help

There are no --api-key, --enc-key, or --credentials flags. Unknown options are rejected.

Agent wallets (x402)

An x402 payment is just a signature, and a wallet key is a vault secret like any other. Store the key at tier 2, have the agent sign the payment payload through vault_exec, and the key is injected into a local signing subprocess — it never enters the model context, and every use needs the owner's approval first (the agent's denied call carries a request id; approval is scoped to that agent + secret, once or for a 15/60-minute window). We ran this end-to-end on Base Sepolia — the verified run is written up at wundervault.com/agent-wallets. Payment-specific policy (spend caps, payee allowlists) is not built yet: compatible, not productized.

Sandbox / demo mode

Set WUNDERVAULT_MOCK=1 to run the server without a wundervault-agent daemon or any credentials. In this mode every tool call returns a representative response clearly labelled [DEMO MODE] instead of contacting the vault — no real secret is ever involved. This exists so you can poke at the tool surface without an account, and so MCP directory scanners and CI (e.g. Glama) can start the server, exercise each tool, and validate the build with no live vault. It is off by default and is never enabled in production.

"env": { "WUNDERVAULT_MOCK": "1" }   // demo/CI only — returns fake, labelled output

Building from source

git clone https://github.com/wundervault/wundervault-mcp.git
cd wundervault-mcp
npm install
npm run build   # compiles TypeScript to dist/
npm test        # run the test suite

License

Licensed under the GNU Affero General Public License v3.0 or later (AGPL-3.0-or-later). See LICENSE.

Wundervault is open-core: this MCP server and the client are open source; the hosted service at wundervault.com is a commercial offering. For commercial or hosting inquiries, get in touch via wundervault.com/contact.