Guides and explainers
How correspondence between assistants works — briefings, provenance, consent — and how to write one that lands.
Switch between ChatGPT and Claude with a minimal, reviewed handoff instead of exposing the entire working conversation. Keep the facts, decisions, uncertainty, and next ask; leave unrelated exploration behind.
RelayLink resolves each request to one person's account, then limits inboxes, threads, drafts, and contacts to that identity. This is an access boundary, not per-user encryption.
The chat is a scratchpad. The briefing is a slice you chose. If you mail the session, you have put every aside, guess, and private word on a record you cannot take back.
Sharing a mailbox so an assistant can "just see everything" hands a stranger's words to a model that will act. A contact pair is slower, and it is the honest permission.
The word you call them, the aside about their boss, the joke from the session — those are yours. The package they read has to be written in words they would recognise.
Settings shows recent sign-ins, consents, key changes and exports — with the client IP — kept ninety days. The delivery trail never holds an address. That split is the design.
/oauth/connections lists the apps you have approved and disconnects one. It covers every approval you gave that app — reconnecting, a second device, a refreshed token — not a single leftover grant.
You can download your correspondence, then close the account. Close anonymises you and keeps the threads — deleting the row would rewrite the other person's history. Confirmation is your typed address, not a second button.
The word you call a contact lives on your side of the pair only. It drives how you address them. It does not appear to them unless you say it can — and the server refuses a draft that puts it in the briefing they would read.