You want Claude and ChatGPT to continue the same conversations, not behave like two people who happen to share contacts. The correct shape is one RelayLink account with two independently removable connections.
Do not register a second RelayLink identity merely because the model or device changed. The account is the person and email address; Claude and ChatGPT are doors into it.
Use one account as the center
The architecture is simple:
Claude ---- its OAuth grant or named key ----\
> one RelayLink account
ChatGPT ---- its OAuth grant or named key ----/
Both connections then see the account's shared inbox, unread state, threads, contacts, private nicknames, drafts, and display identity. A package fetched through one connection is fetched for the account. A contact accepted in the portal is available to both.
That shared state is what lets one assistant continue work another started. It also means the account is not a permission boundary between the two. If Claude and ChatGPT must see different correspondence, they should not both authenticate as this account.
Connect Claude through its own authorization
Open RelayLink's /docs page and copy the MCP address for the running host. Add it as a custom connector in Claude. Claude should send you through RelayLink's browser sign-in and consent:
- enter the emailed RelayLink code;
- check the account address on the consent page;
- approve the Claude application;
- return to Claude and verify the connection with
whoami.
RelayLink always offers OAuth. It is preferable here because the client handles tokens and RelayLink can later revoke every live grant for that application. You do not paste a RelayLink API key into the conversation.
If the particular client cannot complete browser sign-in but can attach an HTTP header, issue a named key from /account/keys in your browser and place it only in the connector's credential field.
Connect ChatGPT to the same account
Add the same RelayLink MCP address in ChatGPT's connector settings. When you add RelayLink as a custom server, enable ChatGPT developer mode so its connector form accepts an arbitrary MCP URL. Complete RelayLink sign-in and consent using the same email account. The application shown on this consent is a separate connection from Claude, even though the RelayLink identity is the same.
Run whoami from ChatGPT as well. Confirm:
- the email is the same one Claude reports;
- the RelayLink host is the intended host;
- the display name is the identity recipients should see.
Do not solve a mismatch by inviting the two accounts as contacts. Stop and reconnect the mistaken assistant to the intended account. Contact pairs represent different people; they are not a synchronization mechanism between models.
If ChatGPT is configured with a key instead, create a key named for ChatGPT rather than copying Claude's key. The same rule applies in reverse.
Keep the two doors distinguishable
Record which authentication method each assistant uses:
Claude: OAuth application
ChatGPT: OAuth application
Framework worker: named API key
This small inventory prevents the common incident mistake of disconnecting an application when the client was actually using a key.
OAuth connections appear under Connected applications. API keys appear under Keys, with their names and last-used times. They are deliberately different controls.
If you approve the same OAuth application on several devices, RelayLink still shows it as one application. Disconnecting it ends every live authorization and token you granted that application, not just one device.
Revoking a named key is narrower: it stops only requests presenting that key. It does not disconnect either OAuth application.
Work across both assistants
Either assistant can check the same inbox and continue the same thread. Before acting on a remembered package or draft, have it read current RelayLink state; the other assistant may have fetched, canceled, or advanced it.
To leave structured context for the other assistant, draft a package to your own RelayLink email and approve it. A self-handoff appears unread in the shared inbox and sends no email notification.
Every assistant-composed outward package still requires your approval. The @relaylink shorthand creates a draft and never delivers by itself. Two connected assistants increase where you can work; they do not weaken the review boundary.
When one connection is no longer wanted, identify its door first. Disconnect its OAuth application or revoke its named key, then verify the other assistant still reports the same account and can read current state.