The vault password is temporarily hardcoded to testing (to be replaced later).
Set BRAVE_SERVICES_KEY before running:
cp .envrc.example .envrc
# edit .envrc and fill in BRAVE_SERVICES_KEYmake run sources .envrc on its own. Alternatively you can use
direnv: run direnv allow once and it loads .envrc
automatically.
make # debug build for the current platform
make run # build, then launch (loads BRAVE_SERVICES_KEY from .envrc)
make frontend # build only the React/Vite frontend
make clean # remove build artifactsOn Linux, run make setup-linux once first to install the GTK/webkit2gtk dev
libraries Tauri needs.
Release bundles:
make mac # .app + .dmg (macOS host)
make windows # NSIS installer (cross-compiled via cargo-xwin)
make linux # .deb + .rpm (via Docker)
make all # all threeGive an agent access to vault items without handing it the vault. The agent
runs brave-vault-agent, which talks to the running app over a local socket.
It never holds the vault secret — only a handle that lets it ask.
brave-vault-agent enroll --label "my agent" # you confirm a popup in the app
brave-vault-agent request --item example.com --action reveal
brave-vault-agent reveal --grant <id> # copies the password to clipboard- Each request pops up in the app for approval, unless you pre-authorized it in the app's Agents panel.
- Access is per-item, single-use, and expires. Revoke any agent or grant from the Agents panel.
A grant authorizes exactly one action, and the token the agent holds cannot be redeemed for the other — you choose which at approval time:
-
revealreturns the plaintext to the agent once (e.g. for the agent to type it). You are opting into the agent seeing the secret. -
useasks the app to perform the login so the agent never sees the password. This is the zero-knowledge path:brave-vault-agent request --item example.com --action use brave-vault-agent use --grant <id> # default: --via inject
--via inject(default) fills the login form via the app's Brave browser extension (seeextension/), driven by a native-messaging host the app provisions automatically. The credential flows app → native host → extension → form and is filled only on the credential's own origin.--via clipboardis an explicit opt-in that copies the password to the clipboard — readable by other local apps, so it is not zero-knowledge and is labeled as such in the audit log.
Same-user caveat: no software boundary fully stops a same-user attacker. use
raises the bar (the secret goes to the browser-spawned host, not the agent), but
true isolation requires running the app/browser under a separate OS user.
For sites that use 2FA (or where you're already logged in), the agent can grab the live session cookies for an approved set of hosts and replay them in its own automation browser — no password, no re-login:
brave-vault-agent request --action session --item "mail.google.com,accounts.google.com"
brave-vault-agent grab-session --grant <id> # prints the cookies as JSON- You approve the exact host set at the popup (a multi-domain login like Google needs its several hosts). The grab covers only those hosts.
- Browser-enforced scope: the extension holds no standing cookie access; it requests permission for exactly the approved hosts at grab time, and Brave prompts and refuses every other origin — so authorizing Google can never touch your bank.
- Bearer-token caveat: grabbed cookies are a full-account session until they expire. The grant is short-lived and single-use; revoke from the Agents panel. Origin-scoping controls which session the agent gets, not what it does with it once it has it.
See docs/agent-access-plan.md for the design.