Most MCP setup guides begin by telling you to generate an API key and paste it into a configuration file. For ChatGPT you can skip that entirely, and you should.
ChatGPT is one of the clients that can complete a browser authorisation. Give it a URL and it does the rest.
The whole setup
Open ChatGPT's settings, add a remote MCP server, and paste:
https://spinrun.ai/mcpThat is it. There is no second field to fill in.
A browser window opens the first time ChatGPT reaches the gateway. You sign in, choose which workspace it may reach, and you are done.
What is actually happening
Worth understanding, because it explains why this is the better path rather than just the shorter one.
ChatGPT's first request arrives without a credential and gets a 401 back. That
response carries a WWW-Authenticate header pointing at a metadata document
describing where to authorise. ChatGPT follows it, runs the OAuth flow, and
retries with a bearer token. Nothing about that required you to know a URL or
copy a value.
The consent screen in the middle does one job: choosing which workspace this client may reach. That is a narrower and more useful question than "do you allow access", and it is the one you actually want to be asked.
Set the rules before you connect anything interesting
ChatGPT with tools is an agent that can act in your accounts. Before you point it at anything you would mind losing, open each connected app and set its policy.
Three rows per app — read, write, destructive — each either allowed or blocked. There is no middle setting, and there is no prompt asking you to approve a call while it is in flight: the gateway speaks stateless HTTP, so a decision has to exist before the call arrives.
That constraint is worth leaning into rather than resenting. "This agent may read anything and delete nothing" is a rule you can reason about calmly on a Tuesday. A dialog box during a long task is a rule you will click through.
Blocking destructive actions everywhere and relaxing where you have a reason is about a minute of work and removes an entire category of bad afternoon.
What ChatGPT can reach
The apps your workspace has connected, as tools named TOOLKIT_VERB_OBJECT —
GMAIL_LIST_MESSAGES, NOTION_CREATE_PAGE, HUBSPOT_UPDATE_DEAL. Blocked
tools never appear in the list at all.
Something worth asking it early:
Look through my unread mail from this week, find anything that reads like a support request, and summarise them grouped by customer.
Two accounts
Connect an app twice and each connection keeps its own credential — a personal
Gmail and a work Gmail stay genuinely separate. Their tools are addressed apart
with a suffix (GMAIL_LIST_MESSAGES__work), while the bare name keeps routing to
your default, so nothing that was working breaks when you add the second.
Next
- Connect an app to Claude — the same flow, other client
- Agent permissions belong at the gateway
- The toolkit catalogue