Agent Web Census

Do remote MCP servers actually work? (Nov 2026)

We sent one MCP initialize handshake to each remote servers in the official MCP Registry checked on 2026-11-19; repeated weekly. 45,311 URLs checked.

56.7%work without auth
33.4%alive but need auth
9.9%failed
293 msmedian initialize latency
7.0median tools per working server
45,311URLs checked

Key facts

Cite this: Agent Web Census, “Do remote MCP servers actually work?”, https://agents.datacensus.org/mcp/health/, check of remote servers in the official MCP Registry checked on 2026-11-19.
Definitions

Outcome of the handshake

Outcome for 45,311 remote MCP servers checked on 2026-11-19Working (initialize OK, no auth)25,702Auth required (alive)15,123HTTP error2,758Timeout92DNS error9TLS error473Connection error1,051Invalid response103
OutcomeServers% of checked
Working (initialize OK, no auth)25,70256.7%
Auth required (alive)15,12333.4%
HTTP error2,7586.1%
Timeout920.2%
DNS error90.0%
TLS error4731.0%
Connection error1,0512.3%
Invalid response1030.2%

Tools per working server

Servers by number of tools (n=24,916)0151-58,5806-106,35611-203,12221-506,06051+783

Median 7.0, mean 16.1; only the first page of tools/list is counted, only for servers that worked without authentication.

Latency

Initialize round trip for 25,702 working servers: median 293 ms, p90 751 ms, measured from one network location including connection setup.

Protocol versions

Version the server answered with when our client requested 2025-06-18. Per the MCP spec a server replies with the requested version if it supports it, so this shows how many servers support 2025-06-18 versus fall back to an older (or newer) version — not each server’s newest supported version.

protocolVersion returnedServers
2025-06-1821,487
2025-03-262,949
2024-11-05656
2025-11-25184
2026-07-2832
2025-11-052

Biggest hosts

HostServers checkedWorking (no auth)
agent-observatory-sensor.nolimit-observatory.workers.dev2,36599.9%
gateway.pipeworx.io1,93696.5%
mcp.apify.com1,2610.6%
api.mcp.ai1,122100.0%
cotlet.dev3980.0%
api.m2mcent.com276100.0%
server.smithery.ai2150.0%
atoms.pipavia.com1450.0%
a2awire.com13599.3%
textforagents.com13175.6%
modevc.com8853.4%
mcp.zovo.one82100.0%
ukdata.reliness.co.uk65100.0%
tooloracle.io5793.0%
api.lazy-mac.com5496.3%
nova.civai.co52100.0%
mcp.jojapi.com4995.9%
thecompanyatlas.com480.0%
bumiatlas.com450.0%
api.meok.ai390.0%

Top 20 hosts by number of registry servers checked; percentage is of that host’s servers.

List of working servers →

Methodology and caveats

For each server in the official MCP Registry that lists at least one remote URL, we test the first remote URL only (identical URLs are checked once) with a single attempt per week: one MCP initialize JSON-RPC POST (protocol 2025-06-18; JSON and SSE responses accepted), 10 s timeout, 256 KB body cap. If it succeeds without authentication we send notifications/initialized and one tools/list to count tools (first page only); no tool is ever called. Legacy SSE-type remotes get a GET and are classed reachable if an event stream starts within 5 s (no tool count). Caveats: a single attempt means transient failures (timeouts, 5xx, rate limits) can be counted as broken; auth_required (HTTP 401/403 or a WWW-Authenticate header) means the endpoint is alive but needs credentials, so we cannot tell whether it works for authorised users; servers listing several remotes are judged by the first only; registry entries can be stale. Checked from a single network location.

Data: health-latest.json · health-servers.json

Last updated: (UTC). Numbers are recomputed automatically from primary sources; see methodology.