SecureVault

Used by your agent. Never shown to it.

How the vault works, and what it does not protect against.

No tool returns a secret

The agent module is an MCP server. Its tools answer "filled", "copied" or "saved", never the value. The vault types a password or code straight into the page in your agents' browser (through DevTools), or puts it on the clipboard for 30 seconds. The value never enters the agent's context or transcript.

Encrypted at rest

Passwords, 2FA secrets, notes, security questions, backup codes, door codes and login guides are encrypted with AES-256-GCM. The key is derived from your master password with PBKDF2-SHA256 (100,000 iterations). Account names, usernames, e-mail addresses, websites, folders and icons are stored in plain text.

You unlock, as often as you chose

Signing in to the SecureVault app with your master password also unlocks the vault for your agents, for as long as the open limit you set in the app says: until midnight by default, or every time, 1 hour, 1 week or 1 month. A folder can have a rule of its own: every time, once a day, a week or a month, or never ask for the entries you tick. The agent module reads these rules and cannot change them.

With the default, the derived key is kept only in memory, in one background process that every agent session shares, and it is dropped at midnight. With a rule of a week or a month, or a never-ask folder, the key of the whole vault is also kept in a private file on the computer where the agent module runs (~/.local/state/secure-vault-mcp/session.json, readable only by you), so a restart does not ask again. The app tells you so when you pick such a rule. The key is dropped, and that file removed, when the last rule has run out, when the vault is locked with the agent module's lock tool (vault_lock), and when you change the master password. Three wrong passwords pause unlocking for five minutes.

Only on the right website

A secret is typed only on the account's own website. For any other site the agent has to say so, and you are asked once in the app; a yes is remembered for that account and site. The page address is checked again right before each value is typed.

You decide per entry

Mark an entry App only and no agent tool reads, fills or copies it. Door codes are always app only. Set an entry to Ask me each time and every use asks you in the app first; one yes covers that entry for five minutes.

Every use leaves a trace

Each use shows a desktop notification and adds a line to an audit log on your computer, with the account name and site, never the secret. Questions to you always come up in the SecureVault app; on approvals, No is the default and Yes needs a mouse click.

Your server, your data

The vault lives in a database on a machine you own. There is no SecureVault cloud: your passwords never pass through a third-party service. Your AI agent can set the server up for you.

The limits

These come from the project's own documentation. Read them before you trust the vault with your accounts.

  • A hostile agent is not made harmless. An agent that controls the same browser could read a filled field back from the page. The guards reduce accidental exposure and wrong-site fills. Keep the vault behind your agent's permission prompt.
  • Some details are plain text. Account names, usernames, e-mail addresses and websites are not encrypted, so anyone who can read your database sees which accounts you have.
  • Your own computer is trusted. Code running as you that can rewrite the agent module or the app, or drive your mouse, is out of scope. Keep those files writable only by you.
  • While unlocked, the app opens without a password on your desktop. Anything that can start the app and read your screen sees an open vault until the open limit has run out: midnight by default, a week or a month if you chose that. Not under the open limit Every time.
  • A long open limit puts the key on disk. With a rule of a week or a month, or a never-ask folder, the key of the whole vault lies in a file on the computer where the agent module runs. Code running as you can read it while it exists, and setting that computer's clock back extends the time it is kept.
  • Folder rules are not separate keys. The agent module enforces them. They decide when an agent must ask you; they are no protection against someone who has the vault key or can write to your database.
  • Desktop app windows are trusted by name. When the vault pastes a login into a desktop app, the app is recognised by its window class, which any program can set for itself.
  • Biometric unlock on desktop keeps its key in a file in the app's documents folder, not in the system keychain. On phones it uses the Keychain or Keystore.
  • Your database is yours to protect. Keep it on your home network or on Tailscale, and never forward its port to the internet: the account names and usernames in it are not encrypted. The server kit's installer limits the port and the database user to private addresses; with Docker the limit is the one address you publish the port on.

Found a problem? Write to josef.nobach@outlook.com.