We check MCP servers and websites from the outside, the way an AI agent sees them, and publish what we find with names and dates. Free tools at x1y1.co/tools.
Useful baseline. On the 1 in 3 that ask for a key: when we looked closely at 61 login-protected vendor MCP servers on November 4, all 61 got the 401 and PKCE right, but 36 didn't advertise Client ID Metadata Documents and 40 didn't advertise the RFC 9207 issuer check. Did you record whether those 401s point to protected resource metadata? That would show how many of the 12,447 a client can actually log in to without manual setup.
Really useful numbers, thanks. Honest answer: no, the report only counted status codes, so I can't split the 12,447 by whether the 401 points to protected resource metadata yet. That's the right next cut and I want it in the next run. Your 61-vendor result (36 without Client ID Metadata Documents, 40 without the RFC 9207 check) shows how big the gap is between "returns a 401" and "a client can log in on its own".
The "1 in 3 won't answer tools/list without a key" number is the quiet killer here. An agent can't make an informed decision to add a server it can't inspect, so those 12k servers are effectively invisible to any automated tool-selection step — they only get picked by a human who already knows what they want. Your friendlier pattern (answer initialize + tools/list for everyone, gate on tools/call) is exactly right, and I'd frame it as a discoverability-vs-registration trade the server author is making whether they realize it or not. One thing I'd love in a follow-up: of the 21k servers that do expose tools, what does the tool-count distribution look like? We've found the context cost of a tool-heavy server is its own hidden tax — a server that dumps 40 tool schemas into the window can cost more in tokens-per-turn than it saves in capability. "Does it answer" is table stakes; "is it worth the context budget" is the next layer of this directory I'd pay attention to.
Thanks Max. The tool-count split is in the full post: 292,447 tools across 20,241 servers, median 7 per server, 815 servers with exactly one tool, 4,325 with 20 or more, and the biggest at 80. So most servers are small, but there is a long tail where the schemas alone take a real slice of the context window every turn. Agree on your framing too: gating tools/list is a discoverability choice, whether the author means it or not.
Separating initialize and tools/list from tools/call auth checks makes a huge difference for anyone wiring MCP servers into automated pipelines. When an agent is evaluating a document-processing or PDF conversion server, hiding tools/list behind a 401 hides basic input constraints too, like whether the server expects a URL, raw bytes, or an out-of-band file upload. Letting clients inspect the schema upfront and gating execution on tools/call lets directories and agents verify compatibility before asking a user to provision a key.
Your 150-tool context math also has an easy-to-miss twin on the response side for data and document servers. Even a compact three-tool PDF server can blow past 50 verbose tool definitions in a single turn if tools/call dumps raw layout coordinates or noisy JSON straight into the chat thread instead of clean Markdown or a resource URI. Keeping the tool list small and returning structured Markdown keeps both the schema overhead and the response payload predictable.
Checking for stale homepage URLs versus actual /mcp endpoints in the registry is another underrated takeaway. Half the broken entries people blame on transport flakiness are just registry metadata pointing at a marketing root path that returns 405 on POST.
Good point on the response side, it gets far less attention than tool-list size. Returning clean Markdown or a resource URI instead of raw layout JSON is the same discipline as keeping the tool list short. And the input-shape point (URL vs raw bytes vs separate upload) is one of the strongest reasons to keep tools/list open and gate only tools/call.
I build Tokenade, which trims what an agent reads, so this one hit home. The useful number is the 12% spam: a registry you can't filter by answering is a list of tools an agent will try and fail on.
Agreed, Paul. And the noise is very concentrated: one repository alone was behind 2,356 entries. Filtering on "does it actually answer" plus "who published it" removes most of it before an agent ever wastes a call.
@nestdaddy thanks for the 2,356 figure. One repository carrying that many entries means a rule on that source removes a whole block at once, not a scatter of single tools.
Yes, and that is how our directory filter works now: one rule on the shared repository, plus one on a description repeated across many entries. Each rule drops a whole cluster at once, and the real servers stay listed.
Building icon and graphics tools for developers since 1990. Makers of IconVectors, IconWorkshop & IconGenerator. Sharing practical SVG tips, icon workflows, and AI-assisted editing.
Good catch: it did not. The scan read only the first tools/list page, and it also kept at most 80 tools per server (361 live servers hit that cap). So the tool totals in the post are a floor, not an exact count. The live, key-required and down numbers are not affected. The scanner now follows nextCursor (up to 10 pages per server) from the next daily run. Thanks for pointing it out.
We're a place where coders share, stay up-to-date and grow their careers.
Useful baseline. On the 1 in 3 that ask for a key: when we looked closely at 61 login-protected vendor MCP servers on November 4, all 61 got the 401 and PKCE right, but 36 didn't advertise Client ID Metadata Documents and 40 didn't advertise the RFC 9207 issuer check. Did you record whether those 401s point to protected resource metadata? That would show how many of the 12,447 a client can actually log in to without manual setup.
Really useful numbers, thanks. Honest answer: no, the report only counted status codes, so I can't split the 12,447 by whether the 401 points to protected resource metadata yet. That's the right next cut and I want it in the next run. Your 61-vendor result (36 without Client ID Metadata Documents, 40 without the RFC 9207 check) shows how big the gap is between "returns a 401" and "a client can log in on its own".
The "1 in 3 won't answer tools/list without a key" number is the quiet killer here. An agent can't make an informed decision to add a server it can't inspect, so those 12k servers are effectively invisible to any automated tool-selection step — they only get picked by a human who already knows what they want. Your friendlier pattern (answer initialize + tools/list for everyone, gate on tools/call) is exactly right, and I'd frame it as a discoverability-vs-registration trade the server author is making whether they realize it or not. One thing I'd love in a follow-up: of the 21k servers that do expose tools, what does the tool-count distribution look like? We've found the context cost of a tool-heavy server is its own hidden tax — a server that dumps 40 tool schemas into the window can cost more in tokens-per-turn than it saves in capability. "Does it answer" is table stakes; "is it worth the context budget" is the next layer of this directory I'd pay attention to.
Thanks Max. The tool-count split is in the full post: 292,447 tools across 20,241 servers, median 7 per server, 815 servers with exactly one tool, 4,325 with 20 or more, and the biggest at 80. So most servers are small, but there is a long tail where the schemas alone take a real slice of the context window every turn. Agree on your framing too: gating tools/list is a discoverability choice, whether the author means it or not.
Separating
initializeandtools/listfromtools/callauth checks makes a huge difference for anyone wiring MCP servers into automated pipelines. When an agent is evaluating a document-processing or PDF conversion server, hidingtools/listbehind a 401 hides basic input constraints too, like whether the server expects a URL, raw bytes, or an out-of-band file upload. Letting clients inspect the schema upfront and gating execution ontools/calllets directories and agents verify compatibility before asking a user to provision a key.Your 150-tool context math also has an easy-to-miss twin on the response side for data and document servers. Even a compact three-tool PDF server can blow past 50 verbose tool definitions in a single turn if
tools/calldumps raw layout coordinates or noisy JSON straight into the chat thread instead of clean Markdown or a resource URI. Keeping the tool list small and returning structured Markdown keeps both the schema overhead and the response payload predictable.Checking for stale homepage URLs versus actual
/mcpendpoints in the registry is another underrated takeaway. Half the broken entries people blame on transport flakiness are just registry metadata pointing at a marketing root path that returns 405 on POST.Good point on the response side, it gets far less attention than tool-list size. Returning clean Markdown or a resource URI instead of raw layout JSON is the same discipline as keeping the tool list short. And the input-shape point (URL vs raw bytes vs separate upload) is one of the strongest reasons to keep tools/list open and gate only tools/call.
I build Tokenade, which trims what an agent reads, so this one hit home. The useful number is the 12% spam: a registry you can't filter by answering is a list of tools an agent will try and fail on.
Agreed, Paul. And the noise is very concentrated: one repository alone was behind 2,356 entries. Filtering on "does it actually answer" plus "who published it" removes most of it before an agent ever wastes a call.
@nestdaddy thanks for the 2,356 figure. One repository carrying that many entries means a rule on that source removes a whole block at once, not a scatter of single tools.
Yes, and that is how our directory filter works now: one rule on the shared repository, plus one on a description repeated across many entries. Each rule drops a whole cluster at once, and the real servers stay listed.
Does your crawler follow nextCursor in tools/list? Otherwise, paginated servers could look smaller than they are.
Good catch: it did not. The scan read only the first tools/list page, and it also kept at most 80 tools per server (361 live servers hit that cap). So the tool totals in the post are a floor, not an exact count. The live, key-required and down numbers are not affected. The scanner now follows nextCursor (up to 10 pages per server) from the next daily run. Thanks for pointing it out.