I like the threat model, but I'd be careful how far the privacy claim goes. Browser-only stops network exfiltration, not a malicious extension or a shared machine reading the DOM, so "nothing leaves your browser" is true in a narrower sense than it sounds. One concrete gap: a JWT decoder that decodes without verifying the signature gives false confidence. Anyone can mint a token with admin claims and a valid-looking payload, and a decode-only view can't tell you it's forged. If you ever add optional signature verification with a pasted secret, that's the feature that would make me actually trust it for staging tokens.