Dray gives Claude every Gmail account you own, then teaches it how you actually write, per person, from the mail you have already sent. Claude drafts the reply. You decide whether it goes.
Claude cannot see how you write. Dray puts that in front of it, inside the tool results, before Claude knows it is about to draft. Then the server rewrites the draft against your full style card and refuses to hand back anything that does not score.
Real tool calls, in the order Claude makes them.
fig. 01 · draft_reply, unedited output shape
Dray measures how you open, how you close, how long your replies run against what they answer, whether you use contractions, whether you hedge, which phrases are yours. It does that separately for every kind of relationship you have, and separately again for the people you write to most.
Then it keeps getting better. Every time you edit a Dray draft before sending, that difference becomes a rule: never opens with “Hope this finds you well”, always cuts the first sentence, signs off shorter to peers than to clients.
Hi Sam, Thanks so much for your patience on this! I'm happy to confirm that Thursday works perfectly on our end. Would it be at all possible to move things to 4pm? I hope this works for you! Looking forward to it! Best, Alvaro
Hey Sam, Thursday works. Can we push it to 4? I'll have the revised numbers with me by then. Does that still leave you enough time before the board pack closes? - Alvaro
fig. 02 · same request, three relationships
Every tool works across all connected accounts. Everything that writes produces a Gmail draft and stops there, so the first time you see an email is before it exists as anything but a draft in your own mailbox.
Bulk changes preview counts and samples and change nothing until you confirm. Mail goes to Trash, never permanent deletion. Dray does not ask Google for the scope that would let it.
In Claude, add a custom connector and paste the Dray URL. No install, no local daemon, no config file to hand edit. It works the same on desktop, mobile and the web.
Google's own consent screen. Finishing it connects your first mailbox in the same step. Add the second and third whenever you like, and they act as one.
Ask Claude to calibrate. Dray reads how you actually write, per relationship, and keeps the measurements. It does not keep the messages.
What Dray keeps is the measurements: how you open, how long you write, the rules learned from your own edits. Exemplars are message ids, and the bodies behind them are fetched from Gmail at the moment a draft is written and thrown away after.
Google refresh tokens are encrypted before they are written down. Every change Dray makes to your mailbox is recorded in an audit log only you can read.
During the beta, Google shows an “unverified app” warning when you connect. That warning is real and we are not going to pretend otherwise. Verification for restricted Gmail scopes takes months and an independent security assessment. We are in the middle of it.
To continue, click Advanced, then Go to Dray (unsafe). If that sentence bothers you, wait for us to finish. That is a completely reasonable thing to do.
Dray’s use of information received from Google APIs adheres to the Google API Services User Data Policy, including the Limited Use requirements.
fig. 03 · the founding hundred is capped by Google, not by us. Unverified apps are limited to 100 users.