State of remote MCP servers — November 2026
- As of 2026-11-19, 56.7% of the 45,311 remote endpoints listed in the official MCP Registry and checked on 2026-11-19 completed an MCP initialize handshake without authentication and 33.4% more answered but demanded credentials, so 90.1% were alive.
- As of 2026-11-19, 14 publisher namespaces with at least 100 endpoints each account for 18.9% of the 45,311 remote endpoints listed in the official MCP Registry and checked on 2026-11-19; their endpoints answered without auth 73.7% of the time versus 52.8% for everyone else.
- As of 2026-11-19, among registry entries outside those bulk namespaces, 63.4% of endpoints first published in 2025-Q3 were alive versus 94.8% of those first published in 2026-Q4 (n=112 and 19,952).
- As of 2026-11-19, across all remote endpoints in the MCP Registry, entries that also list an installable package responded without auth more often than remote-only entries (70.4% vs 55.1%), but outside bulk publishers the gap reverses (48.6% vs 53.1%).
- As of 2026-11-19, 83.6% of working remote MCP servers answered our 2025-06-18 request with that same version, 14.0% fell back to an older one, and the median working server exposed 7 tools.
What we measured
The Agent Web Census sent one MCP initialize request to each remote endpoint in the official MCP Registry. The registry listed 66,932 servers; 45,311 of them have a remote URL and were checked. An endpoint is working if it returned a valid initialize result without authentication, auth required if it answered 401/403 or demanded credentials (alive, but we cannot see inside), and failed otherwise. Overall: 25,702 working (56.7%), 15,123 auth required (33.4%), 4,486 failed (9.9%).
A few publishers dominate the registry
The checked remote endpoints come from 29,565 distinct publisher namespaces (the part of a server name before the slash) and 34,459 distinct hosts, yet the five largest namespaces alone own 15.2% of all endpoints. 14 namespaces list 100 or more endpoints each; together that is 8,560 endpoints (18.9%). Their results swing the headline numbers, and they point in opposite directions: one publisher is almost entirely working, another is almost entirely behind authentication.
| Namespace | Endpoints | % of all | Working | Alive |
|---|---|---|---|---|
| io.github.sadri-dridi | 2,366 | 5.2% | 99.9% | 99.9% |
| io.github.pipeworx-io | 1,936 | 4.3% | 96.5% | 96.5% |
| io.github.mcp-dir | 1,120 | 2.5% | 100.0% | 100.0% |
| io.usefulapi | 1,060 | 2.3% | 0.0% | 100.0% |
| dev.cotlet | 398 | 0.9% | 0.0% | 100.0% |
Working servers are just as concentrated: the five hosts with the most working endpoints serve 22.4% of all 25,702 working servers. Because of this, we report most comparisons below twice: for all endpoints, and for the 36,751 endpoints outside the bulk namespaces (52.8% working, 89.4% alive). Of 34,459 distinct hosts, 52.7% had at least one working endpoint; 33,593 hosts carry a single endpoint.
Newer entries are more likely to be alive
Grouping non-bulk endpoints by the quarter the entry was first published, the share that answered at all rises broadly (not strictly every quarter), from 63.4% for 2025-Q3 to 94.8% for 2026-Q4. The share that works without auth rises from 18.8% to 54.3%. This is a correlation: old entries may have been abandoned, or the registry may simply keep dead entries longer than live ones, and newer entries have had less time to break. We cannot tell these apart from one snapshot.
| First published | Endpoints | Working | Alive |
|---|---|---|---|
| 2025-Q3 | 112 | 18.8% | 63.4% |
| 2025-Q4 | 205 | 21.0% | 56.6% |
| 2026-Q1 | 1,032 | 35.4% | 66.4% |
| 2026-Q2 | 3,590 | 44.9% | 75.0% |
| 2026-Q3 | 11,860 | 54.9% | 87.4% |
| 2026-Q4 | 19,952 | 54.3% | 94.8% |
Registry growth is steep: 38,641 of 45,311 checked endpoints (85.3%) were first published in 2026-Q3 or 2026-Q4.
Where servers run
Outside the bulk namespaces, endpoints on the publisher's own (or otherwise unclassified) domain were alive 90.2% of the time, against 83.1% for endpoints on a shared platform subdomain such as workers.dev, vercel.app or railway.app (n=32,546 and 4,205). The spread between platforms is wide: Vercel 78.7% alive, Render 59.2%, Railway 54.1%, Cloudflare Workers 84.9%, Cloudflare Pages 96.7%. Apify's 1,261 endpoints are all behind auth, so they look 'not working' but are alive. Hosting type is confounded with who publishes and when; we make no claim that a platform causes uptime.
| Host type (non-bulk) | Endpoints | Working | Alive | Median latency |
|---|---|---|---|---|
| Own / other domain | 32,546 | 53.4% | 90.2% | 381 ms |
| Cloudflare Workers (*.workers.dev) | 1,308 | 79.8% | 84.9% | 233 ms |
| Apify (*.apify.com) | 1,261 | 0.6% | 100.0% | n<30 |
| Vercel (*.vercel.app) | 357 | 69.2% | 78.7% | 211 ms |
| Render (*.onrender.com) | 333 | 52.0% | 59.2% | 267 ms |
| Railway (*.railway.app) | 316 | 47.8% | 54.1% | 235 ms |
| Supabase (*.supabase.co) | 144 | 59.0% | 87.5% | 455 ms |
| Google Cloud Run (*.run.app) | 129 | 63.6% | 82.9% | 206 ms |
| Fly.io (*.fly.dev) | 124 | 62.1% | 76.6% | 213 ms |
Working servers answered the handshake in a median 144 ms for bulk publishers and 359 ms for everyone else (p90 301 and 811 ms), measured from a single location.
Packages, repositories and a reversed gap
4,847 endpoints belong to entries that list an installable package as well as the remote (40,464 are remote-only). Across all endpoints the package-plus-remote group works more often (70.4% vs 55.1%). Excluding bulk publishers the order flips (48.6% vs 53.1%), a reminder that composition effects can invert a comparison. Package-backed entries also expose far more tools: median 46 vs 7 for remote-only.
Protocol versions, tools and errors
We requested protocol version 2025-06-18, and servers reply with the version they will use, so this is not a survey of each server's newest supported version. Of 25,702 working servers, 83.6% answered 2025-06-18, 14.0% answered an older version and 0.8% a newer one.
| protocolVersion returned | Servers | Share |
|---|---|---|
| 2025-06-18 | 21,487 | 83.6% |
| 2025-03-26 | 2,949 | 11.5% |
| 2024-11-05 | 656 | 2.6% |
| null | 392 | 1.5% |
| 2025-11-25 | 184 | 0.7% |
| 2026-07-28 | 32 | 0.1% |
| 2025-11-05 | 2 | 0.0% |
Among the 24,916 working servers that returned a tool list, the median is 7 tools and the mean 16.1; 3.1% expose more than 50. The largest is com.agentrxi.mcp/agentrxi with 392 tools, followed by io.github.rajeswaran140/crowvault (379) and com.oneuptime/mcp (358).
Among the 2,758 endpoints that failed with an HTTP status, 404 is the most common at 44.1% (1,217), and 404 plus 410 together account for 1,357 endpoints that no longer appear to exist. 178 endpoints returned 402 Payment Required.
| HTTP status | Endpoints | % of HTTP errors |
|---|---|---|
| 404 | 1,217 | 44.1% |
| 503 | 279 | 10.1% |
| 405 | 246 | 8.9% |
| 502 | 237 | 8.6% |
| 402 | 178 | 6.5% |
| 400 | 172 | 6.2% |
| 410 | 140 | 5.1% |
| 530 | 133 | 4.8% |
Methodology and limitations
One attempt per endpoint on 2026-11-19, from one network location, with a 10-second timeout and one initialize request using protocol version 2025-06-18; only the first remote URL of each entry is checked, and only servers that worked without auth are asked for a tool list (first page only). A single check cannot separate permanent from transient failure. 912 failures (2.0% of all endpoints) carry an error message from the connection path (a tunnel 502) that may stem from the checking environment rather than the server, so failure rates are upper bounds. 'Auth required' servers are alive but unverified. Bulk publisher means a namespace with 100 or more checked endpoints; the threshold is our choice. Registry categories and repository fields are self-reported by publishers. All findings are correlations on one snapshot; the registry is a self-selected population and not a sample of all MCP servers.