ChatGPT MCP: how to add a custom MCP connector

Four things people mix up

MCP (Model Context Protocol)

An open standard for connecting AI assistants to external tools and data. It belongs to no single vendor and does nothing on its own.

A ChatGPT connector

ChatGPT's configuration object for a connection to an outside system. A custom connector is one you add yourself by URL.

The Zylx MCP server

A service Zylx Studio runs at https://zylx.studio/api/mcp. It speaks MCP and advertises tools scoped to one workspace.

A Business Brain

The persistent business context Zylx builds and maintains from a company's connected systems. It is the thing the tools read from — not the protocol, and not the connector.

How ChatGPT's connector model works

OpenAI expects a remote MCP server to answer over streamable HTTP at a stable, publicly reachable HTTPS URL — conventionally one ending in /mcp — and to use OAuth 2.1 when the server needs to know which user is calling. Temporary tunnels and local endpoints are fine for testing and explicitly not acceptable for anything published.

There is a second, narrower path worth knowing about, because it explains why some connectors behave differently: for read-only knowledge sources, OpenAI documents a convention where a server implements standard search and fetch tools so ChatGPT can treat it as a searchable corpus. Zylx does not implement that pair. Its server advertises named business tools instead, which is why Zylx connects as a developer-mode custom connector rather than as a document-search source.

Connecting ChatGPT to your Zylx Business Brain

1. 1

   In ChatGPT, open settings and turn on the developer-mode option for custom MCP connectors. In a Team or Enterprise workspace this is an admin permission, so an owner may need to enable it before it appears for you.

2. 2

   Add a new custom connector and enter the Zylx MCP server URL: https://zylx.studio/api/mcp. There is no separate ChatGPT endpoint — every client uses the same one.

3. 3

   Choose OAuth. Zylx publishes the discovery metadata ChatGPT needs, supports dynamic client registration, and requires PKCE, so the connector can register itself and start the flow without you creating credentials by hand.

4. 4

   Sign in to Zylx and approve the connection on the consent screen. By default this grants an account-scoped, read-capable connection: ChatGPT can list and switch between your own workspaces, and never another account's. Write access is granted only when the write scope is explicitly requested and approved.

5. 5

   Test it. Ask ChatGPT to list my Zylx workspaces, then read my Zylx Business Brain. The first confirms identity and scope; the second confirms the Brain is reachable.

MCP access is included on Pro and above. On a free Zylx plan the connection authenticates and answers identity questions, then returns an upgrade notice for the rest — a deliberate design so a plan limit reads as a plan limit rather than a broken connector.

ChatGPT vs Claude: what differs in practice

Same protocol, same Zylx endpoint, same approval queue. What differs is the surrounding product — and those differences are the whole reason connecting one does not connect the other.

Connecting both is normal and does not duplicate anything: the Brain, the permissions, and the approval queue are shared, so a proposal raised in one assistant is decided in the same queue as one raised in the other. The Claude setup guide covers that side.

What changes once ChatGPT is connected

• Ask questions against real business context instead of re-explaining the business each session — what changed, what is underperforming, where revenue actually comes from.
• Trace an answer to its source. Read tools return the record behind a number and how fresh it is, so a claim can be checked rather than trusted.
• Queue work without executing it. Proposals land in the Zylx approval queue and wait for a human, which is the default behaviour rather than a setting.
• Switch between brands in one conversation when the connection is account-scoped, without a second connector.

Limitations

• Custom MCP connectors in ChatGPT are a beta capability behind a developer-mode setting. Availability varies by plan, workspace policy, and rollout — Zylx cannot change that.
• Zylx does not implement the standard search and fetch tool pair, so it connects as a custom connector rather than as a document-search knowledge source.
• An assistant can only reach the workspace its token resolves to. A caller-supplied workspace ID that does not match is refused, not silently honoured.
• Write-class tools refuse a read-only token and return the exact missing scope instead of failing quietly.
• Approval tools refuse a machine identity — an assistant cannot approve its own proposal.
• MCP access is a paid capability. On the free plan the server answers session tools (who am I, which workspaces) and returns an upgrade notice for everything else.
• Tokens expire after 90 days. Regenerating one disconnects any assistant still using the old value.

ChatGPT MCP FAQ