Who the Claude connectors serve: a crowded directory, nearly empty from the customer's side

On 2026-09-04 the claude.ai connector directory held 2,383 connectors, up from 2,105 on 24 August. My agents keep that count, and every time it grows I get the same question from people who run a practice, a clinic or a firm: is my shelf already taken? The counter cannot answer that. It counts doors, not who is allowed through them.

So on the same day I walked three categories whole, health, life sciences and legal, and my agents labelled every card by who it serves. Between them the three hold 170 cards. 18 face a business’s own customers. Of those, 6 belong to a business doing a service for named clients rather than to a software product whose users are its customers, and exactly one is a single local venue. The rest serve people inside companies, serve engineers, or serve the market at large, and 5 cards do not say enough to be labelled at all.

What the labels mean

Before the table, the words in it. Each label describes the person on the other end of the connector, read from nothing but the card’s own name and description.

• customer: serves the business’s own customers. A studio answering about its own classes, an app letting its users read their own data.
• staff: serves people working inside a company. A contract tool for the legal team, a lab notebook for the researchers.
• developer: serves engineers.
• platform: serves many businesses at once, or is a public reference source. A court-records search, a literature database.
• unclear: the card does not say. This is reported as its own number and never folded into another bucket.

Category	Walked	Customers	Staff	Developers	Platforms	Unclear
Health	27 of 27 · whole	12	2	0	10	3
Life sciences	44 of 44 · whole	1	19	5	18	1
Legal	99 of 99 · whole	5	62	1	30	1
170		18	83	6	58	5

Three categories, whole

These are complete walks, not samples, so the wording in the table is plain: in the category there are this many. Three things the table shows and one it does not.

Health is the outlier, and the reason is that it is full of consumer apps. Most of its customer-facing cards are a fitness tracker, a nutrition log, a home sensor: products whose users are their customers. That is a real category of connector, and it is not the thing a clinic is asking about.

Legal is the largest of the three and the emptiest in the way that matters. Its customer-facing cards are legal services that take work in and hand it back, or self-serve products for people without a lawyer. Not one is a law firm answering its own clients.

Life sciences is built for people who work in labs and for the databases they query. One card in the whole category faces a customer.

The thing the table does not show is the second reading, so here it is by name. The cards that are a business answering its own clients: a Toronto fitness studio, a mobility-equipment rental, a test-results service, a company-incorporation service, a contract-review service, and one large consumer legal brand. One of those is a single local venue. Every one of these judgements sits in the dataset as a flag on the card, so you can disagree with it card by card instead of with me in general.

What the two sides look like

The difference is not the industry and not the technology. It is who is holding the phone.

On the customer side: a fitness app whose connector lets its users analyse their own workout data. A test-results service whose connector lets a customer read their own results. And the studio, whose connector answers about its own classes, schedule, pricing and instructors, straight from its own published pages.

On the staff side: a contract platform whose connector lets a company’s own team search its contracts in plain language. A connector giving natural-language access to a laboratory’s electronic notebook. A data-room product whose connector lets a deal team manage its own room from the chat.

Same directory, same category page, same install button. The person the tool is built for is the entire difference, and the card usually tells you in its first line.

Your shelf

If you run a law firm: there are 99 cards in legal, and none of them is a law firm talking to its own clients.

If you run a clinic, a physiotherapy practice or a diagnostics lab that sees patients: the customer-facing cards in health are, nearly all, apps serving their own users. The one card that reads like a neighbourhood business is a fitness studio.

If you run anything in life sciences that has customers rather than users: one card in the category.

That is what the counter hides. A number like 2,383 reads as a crowded market. Stand on the customer’s side of the counter, in one of these three categories, and you can count what is there on one hand.

How the count was made

My agents run the census. The directory shows its cards only to a logged-in profile, and Anthropic’s terms rule out automating that page, so the one step a person does is open each of the three categories in a browser and scroll it to the end. A snippet the agents wrote reads every card as the directory rendered it: name, description, and whether it sits in the community section. Nothing else on the site was touched. Everything after that, the dataset, the labels, the table, is their work.

The labels were assigned from the card’s own name and description only, under the rule above. Where a card does not say who it serves, the label is unclear, and unclear stays unclear. I set the rule and I stand behind the count.

Three boundaries a reader should know. A label reads the card, not the product behind it: a connector whose card undersells itself gets the label its card earns. A connector can sit in more than one category, so 170 is the number of cards inspected and 168 the number of distinct names; two names appear twice. And a whole walk is a whole walk on the day. The directory changes, which is why the number carries a date and why the next pass gets an ordinal rather than a promise.

The dataset

Everything above renders from one file: /data/shelf-map.json, CC BY 4.0. It holds the labelling rule, every card and every label. A reader who thinks a label is wrong can name the card and say why, which is the point of publishing the labels rather than the totals.

Changelog

• 2026-09-11: wording. The count is run by my agents; a person only opens the categories in a logged-in browser. The earlier “by hand” undersold the method.
• 2026-09-04: pass #1. Health, life sciences and legal walked whole on a logged-in claude.ai profile: 170 cards captured, labelled and published. Directory total on the same day: 2,383.