Connectors

Connectors are the escape hatch from "one assistant, our engines". There are two kinds and they do genuinely different things.

Tool connectors (MCP)

An MCP connector points at a server that exposes tools — a search index, a calendar, an internal API, a database. When one is switched on for a chat, the assistant can call those tools mid-answer and use what comes back.

While a tool call is running you will see what it is doing. That status disappears as soon as the answer starts arriving, because by then what a tool was doing a moment ago is no longer what you are waiting on.

Tool calling is capped at a few rounds per message. A model that keeps deciding it needs one more lookup would otherwise loop.

Your own model key (BYOK)

A BYOK connector answers with a model you pay for, using a key you provide. That chat lives on its own route and never touches the shared engine — no racing, no fallback, no quota borrowed from anyone else. If your key stops working, that chat stops working; it does not quietly fall back to ours.

This is the one place in the product where a provider is named, and only because you named it. You chose it and typed it in.

Keys are encrypted before they are stored and are never returned by the API — not to you, not to an administrator, not in an error message.

Availability

Connectors are a Pro plan feature. See Account and billing for what that currently means, which is not what you might expect.

The gate is enforced on the server, not by hiding the button. A request that carries connectors without the plan is refused with a message saying so — it is not silently run without them, because that would leave you believing a tool had been consulted when nothing was.

Limits and safety

• Connectors cannot be pointed at private or internal network addresses. A server

that would proxy requests into its own network on request is a security hole, not a feature.

• There is a cap on how many connectors one account can add.
• A tool result is data, not instruction. Treat what a connector returns the way you

would treat any other untrusted input — and be deliberate about which tools you switch on for a conversation, because the assistant will use them.