A few days ago I learned something about Claude that I should have known months ago.
I run about fifteen different integrations across Claude's ecosystem. Cloudflare, Google Drive, Gmail, Stripe, a handful of specialty tools for legal research and local SEO, a few custom MCP servers I built myself. Normal stuff for anyone running a serious AI-powered workflow.
I assumed — because the documentation strongly implies it, and because it's how every other piece of software I've ever used works — that these integrations only activated when I used them. Click the connector, the tool wakes up. Close the conversation, the tool goes quiet. Clean separation.
I was wrong. They were all loading into every conversation I had, all the time, regardless of whether I was using them.
And here's the part that really stung: some of them I thought I had disconnected. I'd toggled them off in a chat interface and moved on, genuinely believing they were gone. They weren't. They were still authenticated, still loading their tool definitions into my context window, still quietly taxing every session.
If you use Claude seriously — if you've added connectors to your account over the past year or two — there's a very good chance the same thing is happening to you right now. Let me explain what's going on.
The Three Layers
Claude extensions have three distinct states, and the platform does not explain this anywhere. Most of the confusion — including mine — comes from conflating them.
Installed is the top layer. A connector exists in your Settings → Connectors list. It's been registered to your account. It hasn't necessarily done anything yet.
Connected is where the cost lives. You've completed the OAuth handshake. The connector can now reach its service on your behalf. And — this is the critical part — its tool definitions, its system instructions, and its capabilities are now loaded into every single conversation you start. New chat in a completely unrelated project? Doesn't matter. Still loaded.
Enabled in this conversation is the third layer and the one that fools everyone. It's a toggle in the "+" menu of your current chat. Flipping it off makes Claude stop calling that connector's tools. But the tool definitions are still in your context. The connector is still "Connected" at the account level. The toggle is cosmetic — a per-conversation mute, not a disconnection.
This matters because when most people think "I turned that off," they're thinking of the cosmetic toggle. The actual fix lives one layer up, in account settings, behind a button explicitly labeled Disconnect.
Why This Is Happening To Everyone
Here's the architectural reality: Projects in Claude do not scope integrations. At all.
Projects are very good at scoping content. Documents you upload to Project A don't leak into Project B. Chat history is isolated. This separation is real and well-implemented.
But integrations — connectors, MCP servers, skills — live one level up, at your account. They load globally. Every project, every conversation, every new chat. Your Legal Research connector is sitting in your context when you're asking Claude about recipes. Your Stripe connector is loaded when you're drafting an essay. Whatever you've connected, you're paying for it in every conversation, whether you use it or not.
This is a known gap. The Anthropic GitHub issue tracker has active feature requests asking for per-project connector configuration. As of this writing, nothing has shipped. Users are left doing manual account-level hygiene.
My Audit, In Specifics
When I finally sat down to audit my own account, here's what I found.
Four connectors I had not used in over a month — one tied to a business vertical I'd explicitly deferred, one a legacy experiment I'd forgotten about, one I couldn't even remember adding. All four were actively loading into every conversation. Combined context tax: a few thousand tokens minimum, plus whatever latency each connector added at session startup.
Three connectors I thought I had disconnected. I remembered clicking something, I remembered feeling like I'd taken care of it. When I checked Settings, they were all still there with a Disconnect button waiting to be pressed. What I'd actually done, months earlier, was toggle them off in the "+" menu of a conversation I no longer had open.
Two connectors I genuinely couldn't assess. A local SEO tool and a model repository. Was I using them? Sometimes? Probably? The fact that I couldn't answer the question was itself the answer.
Total: nine connected integrations. Four immediate disconnects. Two defer-until-proven-needed. Three genuine keepers.
I disconnected the four. Claude felt faster the next session. That might be placebo. It might also be a few thousand tokens and a few hundred milliseconds of MCP handshakes I was no longer paying for.
What You Can Actually Do About It
Here's the practical audit, in the order I'd run it:
- Open a fresh browser tab to your connector settings. Either
claude.ai/settings/connectorsorclaude.ai/customize/connectors— both paths lead to the same list. Look at every connector that shows a Disconnect button. - Ask yourself a simple question for each. When did I last actually use this? Not "when did I last see it in a menu," but when did Claude actually call one of its tools for me. If you can't remember, that's data.
- Disconnect anything on the fence. The re-authentication flow for most connectors takes about ten seconds. You can always reconnect. You're not burning bridges. You're just cleaning up.
- Start a new chat to verify. Your current conversation has the old connector list baked in from session start. Changes you make in settings don't affect in-flight chats. A fresh chat shows you the new reality.
- Commit to revisiting the list monthly. Connectors accumulate the way browser tabs accumulate. One a week adds up. A thirty-second monthly audit beats a mysterious slowdown three months from now.
The Honest Part
I want to be clear about something: this isn't a complaint about Anthropic.
They're building a capability at a pace that is genuinely unprecedented. Connectors launched last summer. Skills followed. MCP evolved in parallel. Cowork came in January. Each piece was shipped because it unlocked real value for real users — and each one, at the moment of launch, did not have a perfectly integrated governance story. That story is being retrofitted now.
This is what "building the plane while it's flying" actually looks like from a user's perspective. You get capabilities that didn't exist six months ago, but you also get surfaces that don't quite explain themselves, edges that don't quite meet, and the occasional moment where you realize you've been operating on a mental model that wasn't quite right. That's the tradeoff for being on the bleeding edge. If you wanted polished, documented, and stable, you'd be using something five years old and half as capable.
The mistake isn't that Anthropic shipped fast. The mistake would be assuming the fast-shipped thing works the way you intuited it works. It probably doesn't. Not yet. Check your assumptions periodically and you'll save yourself a lot of mysterious frustration.
What Better Would Look Like
Since we're here, let me say what I'd love to see — both as a user and as someone who thinks about these problems for a living.
A unified view of every extension loaded into a conversation, with actual context-cost estimates. You have twelve thousand tokens of overhead before you've typed a word. You might be fine with that. You might not. The point is knowing.
Usage telemetry that tells me when I last called a tool. My email client knows when I last opened a thread. Why doesn't my AI platform know when I last invoked a connector?
Per-project allowlists. I maintain multiple distinct workstreams and each one needs a different toolkit. Let me declare the toolkit per project. Don't make me remember to toggle.
Proactive nudges. "You haven't called this connector in forty-seven days. Disconnect?" This is not a hard problem. My calendar does this for subscriptions. My bank does this for recurring charges. My AI tools could do this for extensions.
None of this exists yet. Some of it probably will, in the next six to eighteen months. Some of it is tooling a builder could produce themselves if they wanted to. In the meantime, the hygiene is manual. Monthly audit. Disconnect decisively. Reconnect cheaply when needed. That's the game.
The Meta Lesson
The deeper takeaway, for anyone building their own workflow around AI right now:
Friction is your teacher. Every moment you hit an edge — a connector that shouldn't have been there, a setting that doesn't do what you thought, a behavior that caught you off guard — that friction is pointing at something real. Either the platform has a gap, or your mental model does, or both. The work is figuring out which.
The people who will look back in two years and say "I wish I'd gotten serious about AI earlier" are the people who are currently waiting for a polished, documented, stable system. There isn't one coming. The polished version will exist, eventually, but by then the people who stumbled through the messy version will have a ten-thousand-hour head start on how to actually think with these tools.
I'd rather stumble now. Even when it means discovering I'd been carrying nine suitcases into every conversation and only actually needed five.