222 answered openly
These servers list their tools to any caller. For discovery that is a feature, not a flaw: the tool list is the connector's shop window, and these publishers treat it that way.
Original research · August 2026
The State of Claude Connectors
We catalogued every listing in the Claude connectors directory (1,253 listings, 15 categories, 1,206 publishers) and then did the thing directories never do: called every published endpoint to see what the servers actually serve. Here is what the full population shows.
Finding 01
24,660 tools are published. We checked how many are really there.
Every connector declares a tool count in its listing, and almost nobody ever tests those claims against the server behind them. We did, for all of them.
24,660tools published across 1,253 connectors
222endpoints answered and listed their tools
3,072tools enumerated first-hand
31%of those published a wrong count
Of the 222 connectors whose servers answered, 68 publish a tool count that does not match their own server, 56 of them understating it. Listings drift as servers evolve, and nothing in the marketplace process catches it. The widest gap is Deel: listed at 95 tools, its server answers with 449.
The full inventory of every verified tool is published at verified MCP tools, and wherever a server answered, the connector's profile carries the measured tool list. For the credential-gated majority, the counts shown are what the listings publish.
Finding 02
What happens when you actually knock
Calling 1,253 published endpoints with a clean, unauthenticated MCP handshake sorts the directory into postures no listing page shows.
836 require credentials first
The dominant posture: the server exists and speaks MCP, but will not enumerate tools without an authenticated session. A deliberate design choice, and the majority of the directory.
19 are blocked at the CDN edge
Not an auth gate: a bot filter in front of the endpoint refuses the connection outright. Distinct from a credential gate, and worth distinguishing: it can also block legitimate clients.
176 refuse in other ways
Unspecified errors, IP allowlists, and failed handshakes. Combined with the 48 listings that publish no endpoint URL at all, a meaningful slice of the directory cannot be evaluated from the outside.
Finding 03
Category coverage is extremely uneven
The five largest categories hold 80% of everything listed, and the taxonomy is loose at row level, since publishers file their own listings.
Finding 04
Verification is the exception, not the rule
61.5% are community-listed
770 of 1,253 connectors are self-published with no review step. The tier says nothing about quality either way, but it does mean most of the directory has never been looked at.
37.8% carry partner verification
474 connectors went through review. Regulated verticals skew heavily here; finance carries a far higher partner share than the directory average.
9 were built by Anthropic itself
Everything else is third-party. For a publisher, that means the field is open: verification is a differentiator precisely because so few have it.
The median connector publishes 11 tools
Against a mean of 19.7 and a long right tail reaching 255. Depth varies far more than presence does.
Finding 05
Almost nobody has built a second one
97.7% of the 1,206 publishers with a listing have exactly one connector. The directory is nearly all tail; there is no established head of multi-connector publishers to compete with, which is the clearest indication of how early this market is.
And only 418 companies appear in both the Claude and ChatGPT directories, about 35% of this one. Since a single MCP server can serve both marketplaces, the other ~65% are one distribution decision away from doubling their reach. The unified AI connectors directory shows exactly who ships where.
Method & reuse
How this was collected, and what you may do with it
The Claude connectors directory was walked in August 2026, capturing 1,253 listings, the full listed population rather than a sample. Every published endpoint then received an MCP initialize handshake and a single tools/list request from a clearly identified Node8 user agent, with no credentials and no tool invocations. Refusals were recorded as results, never worked around.
The compilation, measurements, and analysis are published under CC BY 4.0: quote, republish, and build on them, including commercially, with attribution and a link to this page. Suggested citation: Node8, The State of Claude Connectors, August 2026. https://node8.ai/state-of-claude-connectors
Figures are accurate as of August 2026 and will drift as the directory changes. Companion research on the other marketplace: The State of ChatGPT Apps.
FAQ
Research questions, answered
How many connectors are in the Claude connectors directory?
1,253 listings were live when the directory was captured in August 2026, resolving to 1,253 connectors after collapsing re-listed duplicates, from 1,206 distinct publishers across 15 categories.
Are the published tool counts accurate?
Often not. Node8 called every published endpoint directly; 222 answered with their tool list, and 31% of those publish a count that does not match their own server, usually understating it. The widest gap found: a connector listed at 95 tools whose server returned 449.
How were the endpoints tested?
With the minimum possible: an MCP initialize handshake followed by a single tools/list request, from a clearly identified Node8 user agent, with no credentials, and never invoking any tool. A refusal was recorded as a result, not worked around. Servers that answered are quoted exactly as they responded.
What do the verification tiers mean?
770 connectors are community-listed (61.5%), meaning self-published. 474 carry partner verification (37.8%), which is a review step. 9 were built by Anthropic itself. The tier says who published and whether it was reviewed, not whether the connector is good.
How much overlap is there with the ChatGPT app directory?
418 companies appear in both directories, covering 440 Claude connectors, about 35% of this one. Since one MCP server can serve both marketplaces, the gap is a distribution decision, not a technical constraint.
Can I reuse this data?
Yes. The enumeration, measurements, and analysis are published under CC BY 4.0, so you may quote, republish, and build on them, including commercially, provided you attribute Node8 and link to this page. The underlying listing facts come from Anthropic’s public directory and each publisher’s own public endpoint, and are not ours to license.
The directory is 61.5% unreviewed and nearly all single-connector publishers.
That is what an open field looks like. Node8 designs the tool surface, builds and secures the MCP server, and takes it through review on Claude, ChatGPT, and Copilot. Most first connectors ship in eight to twelve weeks.