Four Connectors, Not Forty

The problem stopped being availability

Two years ago the question was whether Claude could reach your tools at all. That question is dead. Depending on who's counting and what they count, Claude's connector directory holds somewhere between 369 and 439 integrations across as many as 30 categories, and one well-regarded guide from June puts the number of polished, one-click directory apps at closer to 50. Those numbers disagree because they're measuring different things: the curated directory, the full browsable list including third-party servers, and the subset that's genuinely one click.

The disagreement is the point. Anthropic doesn't publish a headline number at all, so every figure you'll see quoted is somebody else's count of somebody else's list. Nobody can tell you exactly how many connectors exist. What everyone agrees on is that there are far more than you need, and no obvious way to tell which ones matter for you.

So users do one of two things. They connect nothing, and Claude stays a very good writing assistant that knows nothing about their business. Or they connect everything, and Claude becomes slower, vaguer, and weirdly bad at choosing the right tool. Both feel like the platform's fault. Neither is.

Here's the reframe that fixes it. Stop asking which connectors are good. Start asking what shape a working setup has. Because every genuinely useful Claude workflow, in any business, has the same shape, and once you can see it you can build your own stack in an afternoon.

Every working setup is the same four links

Watch someone use Claude well and you'll see four moves, always in the same order. Raw material comes in from somewhere. Something applies judgment to it. Something builds the artifact. The artifact lands where it has to land to count as done.

Now the diagnosis. Go look at your own connector list. Most users have four or five things plugged into the Source link, nothing at all in Judgment, and nothing in Land.

That's why their AI work dies in a chat window. They can ask Claude anything about their business and get a smart answer, and then they copy the answer out by hand, paste it somewhere, and reformat it. The chain has a beginning and no end.

The Judgment link is the one nobody connects

A connector moves data. That's its whole job, and it's an honest job. The Shopify connector can hand Claude your entire product catalog in a second. It has no opinion whatsoever about whether your product descriptions are any good.

That gap is where most disappointment with AI tools actually comes from. Users connect their data, ask for something, and get a competent, generic answer. The data was there. The judgment wasn't.

Judgment is what a skill pack supplies. A skill is a folder of instructions that gives Claude a specific way of working: what to ask before it writes, what a good version of this document looks like, what the common mistakes are, what to refuse to do. A plugin bundles a set of related skills together, along with the connectors they expect and any background agents that come with them.

The difference in output is not subtle. Ask Claude cold for a client proposal and you get a competent template. Ask through a delivery-focused skill pack and it asks about your rates first, warns you which parts of the scope are likely to cost you money, writes the proposal, writes the covering email, writes the follow-up sequence, and tells you honestly if a proposal is even the right next move.

Where skills and plugins actually run

This matters more than it looks, because the answer differs by surface and by plan. Here's the current position.

The line that catches people out is the last one. A plugin's skills work everywhere, but the background agents bundled with it only run in Cowork. If a pack advertises a weekly sweep that flags your overdue invoices, you get that sweep on demand in chat and automatically in Cowork. Same intelligence, different trigger.

Why more connectors makes Claude worse

This is the part nobody selling you an AI course wants to say. Every connector you enable costs something, and the cost is paid out of the same budget that holds your actual conversation.

When a connector is active, the descriptions of everything it can do get loaded so Claude knows what's available. A single connector might expose twenty or thirty separate actions, each with its own description. Connect eight of them and a meaningful slice of the working memory is spent before you've typed a word.

One engineering team measured this in a development setup and found that three fairly standard integrations consumed over 20 percent of the available context before the assistant started working. That was a code-heavy setup, which is heavier than most business chat, so treat it as an illustration rather than a law. The direction is what matters.

The symptoms are the frustrating kind, because nothing visibly breaks. Instructions get ignored. Answers drift generic. Claude reaches for the wrong tool, or asks you something it should already know. It feels like the model got worse. It didn't. It got crowded.

The four-connector rule

One connector per link. Four total. When you want a fifth, it replaces one rather than joining the set.

This sounds arbitrary and isn't. It's just the observation that a workflow with two Sources usually means you haven't decided where your real data lives, and a workflow with two Lands usually means you haven't decided what finished looks like. The constraint forces the decision.

You'll break it eventually, and that's fine. A seller running two storefronts genuinely needs two Sources. But break it deliberately, after the four-link version works, not by accident on your first afternoon of enthusiasm.

Read or write decides what a connector can be

Here's the single most useful thing to know before you connect anything, and almost nobody says it out loud: most connectors are read-only.

Ahrefs exposes 61 tools through Claude and every one of them reads. Semrush has 14, all read. Helium 10 has 34, all read, and its own documentation is explicit that it cannot place a bid, pause a campaign, or edit a listing. Triple Whale states plainly that it cannot mutate your business data, ad accounts, campaigns, or budgets. None of that is a flaw. Reading is what those tools are for.

But it settles where each one can sit in your chain. A Source can be read-only. A Land cannot. If the only connectors you have are read-only, you don't have a chain, you have a very well-informed conversation that still ends with you copying text into another window.

They don't fire on their own

A connector is not an automation. It does nothing until you start a conversation and ask. Buffer's official server, released in May 2026, is a good example: Claude can list your channels, see your queue, draft posts, and schedule them. What it cannot do is watch anything or react to an event. Buffer's own documentation is clear that the server only works inside a conversation you start. Gorgias says the same about its helpdesk connector: it will answer a ticket when you ask, and it will never notice one arriving.

If you want something to happen at 7am on Monday whether or not you're at your desk, a connector alone won't do it. That's what Cowork's scheduled tasks and background agents are for, and it's the strongest practical reason to care which surface you're working on.

Most writes land as drafts, but check which

Kajabi's server, included on every Kajabi plan at no extra cost, can build pages, write emails, and set up offers. Kajabi's marketing page says every write action lands as a draft. Kajabi's own technical documentation says otherwise, and the documentation is right: creating a page, course, or email does produce a draft, but community posts, direct messages, pinned announcements, navigation links, and theme edits all take effect immediately. On theme content the docs state it flatly, that changes take effect on published pages and there is no draft layer.

Mailchimp goes the other way and will draft a campaign but refuses to send it, so a human has to press the button. Typeform's connector will publish a form live and its documentation warns you to check it first. There's no universal rule here. Read the two paragraphs of documentation before you let a connector write anything.

The connectors that aren't what their name says

This section will save you more time than any other. A connector's name tells you which company built it. It does not tell you what it does, and in several cases the gap is wide enough to waste an afternoon.

Five stacks you can copy today

Each stack fills all four links with tools built for that trade, names the pack that supplies the judgment, and lists the specific skills the chain runs through. Find your row, then read that one card.

Build your own stack in twenty minutes

None of those five is your business exactly. That's fine, because the method is the transferable part, and it runs backwards from the way most people approach it.

Don't start with the directory. Start with the deliverable you resent making.

Step six is the one people skip, and skipping it is why so many carefully built setups get abandoned. You cannot tell whether a stack works by looking at it. You find out on the Friday when you're tired and you'd rather not.

What to do this week

Open your connector settings and count what's there. Sort each one into Source, Judgment, Make, or Land. Most users find three or four Sources, an empty Land, and a couple of things they connected months ago out of curiosity and have never used once.

Turn off everything you can't attach to a job you actually do. This will feel like losing capability. It isn't. Every connector you remove gives Claude back room to think about the ones you kept.

Then fill the empty link. If it's Land, that's a fifteen-minute fix and the largest single improvement available to you, because it's the difference between advice and finished work. Check the tool list writes before you commit. If the empty link is Judgment, install the pack that covers your trade and run one job through it.

The users getting real work out of Claude right now are not the ones with the most connectors. In most cases they have noticeably fewer than the people struggling. What they have is four that form a chain, and a clear idea of which weekly job that chain exists to kill.

Pick your job. Build the four links. Run it once this Friday.