Connecting your mail to an assistant means the model sees whatever is in those messages. Login and verification emails are a good example.
A one-time code is meant for you, once, for a minute, and then never again. Nobody sets out to hand a bank's verification code to an LLM. You just ask it to check your mail, and in comes the code, riding the same wave as the receipts and the reminders and the promotional clutter, perfectly ordinary and entirely exposed.
Redaction for Gmail
So this week I added security-code handling to the Gmail toolkit, and it's on by default. When the toolkit reads a message, it detects login codes, one-time passwords, password reset URLs, and magic links, and replaces them with a [REDACTED] placeholder before the content is returned to your assistant. The rest of the message comes through as normal.
It works off a dynamic list of known security-code senders combined with text patterns for the codes and links themselves. The point is to protect you if an assistant, a local agent, or some connected app is ever compromised, and to stop a malicious email from using prompt injection to talk your assistant into handing over a code.
If you want something other than the default, the toolkit's settings let you switch from redacting just the codes to withholding any message that looks like it contains one, or turn the handling off entirely. The setting applies across all of your sessions. I'd leave it on.
You can read more in the Gmail section of the docs.
A MusicBrainz toolkit
The music side of Toolforest already had ListenBrainz, Last.fm, and Apple Music. This week I added a MusicBrainz toolkit to sit alongside them. Toolforest has kept a replicated MusicBrainz database for a while to back the ListenBrainz toolkit, and the new toolkit puts that data directly in your assistant's hands.
The toolkit resolves any artist, recording, or release to its canonical MusicBrainz identifier and reads the credits behind a track: who produced it, who engineered and mixed it, and who played what. Credits in MusicBrainz are sometimes attached to the album rather than the individual song, so the toolkit also checks the release level. A question like "who produced this track" still gets an answer even when the credit lives one level up.
There are also relationship tools, so you can start from a band and branch outward to its members and the other projects they belong to. Put together with ListenBrainz, that means your assistant can take your most-played tracks and tell you the recurring producers behind them, or map the collaboration web around a band you like. There are now examples showing both.
As always, ideas and suggestions are welcome at gerrit@toolforest.io.