Why "Connect Anything" Isn't on the Roadmap Here
A lot of agent and automation tools compete on connecting to any API you point them at. Here's why the Connectors page deliberately doesn't, and why that's a defensible trade-off rather than a missing feature.
"Connect to anything" is a common competitive claim among agent and automation tools — a form where you paste a base URL and a key, and the tool figures out the rest. It's an easy feature to demo and a genuinely risky one to ship: an agent given an arbitrary base URL and an arbitrary credential is, functionally, server-side request forgery with a credential attached. A model deciding where to send an authenticated request, without anyone having reviewed what that endpoint can do, is exactly the kind of open-ended surface that named 2026 agent-security writeups keep flagging.
The deliberate alternative
The Connectors page in Settings doesn't offer that form. It offers a reviewed catalogue instead — a fixed list of third-party systems, each with a base URL that's a constant in the platform's own code and a fixed set of read-only endpoints. You provide a key for something already on that list; you don't define a new destination for the agent to reach.
That's a narrower promise than "connect anything," and it's narrower on purpose. Every entry on the list is something that's actually been reviewed — the endpoints it exposes, what it can and can't do, whether "read-only" is actually enforced rather than assumed.
The same three rules that govern AI Keys, applied here too
This page was built to the same three commitments as the AI provider key page: never show a stored key back to you, never invent a connector the backend doesn't actually serve, and never offer a form the deployment can't honor. Copying that discipline from a page that already had to solve the "credential handling done right" problem was a deliberate choice, not a coincidence.
Why this is a trade-off worth explaining, not apologizing for
A narrower, reviewed catalogue is a genuinely different product decision than "integrate with anything," and it's worth stating plainly rather than treating it as a gap to eventually close. The absence of an open-ended connector form isn't an unfinished feature — it's the position: an agent's reach into your other systems should be a reviewed, bounded list, not a form that trusts whatever URL gets typed into it.