Live · updated daily

MCP servers statistics

What our own probes see across 61,862 MCP servers: how many answer, which protocol versions they speak, how ready they are for an agent to call, and what they cover. All protocols

As of 2026-11-12 01:39 UTC Methodology JSON CSV Cite

How many agent tools actually work, and why the rest do not

Indexed
61,862
· 0 vs 7d ago
Listings in our index
Reachable
53,021
▼ 144 vs 7d ago
85.7% of indexed · Answered the last probe
Protocol verified
35,921
▲ 118 vs 7d ago
58.1% of indexed · Full MCP handshake
Readiness A or B
34,478
▲ 852 vs 7d ago
55.7% of indexed · Ready for an agent to call
01 · Funnel

85.7% of indexed MCP servers answer our probe, and 58.1% complete a full handshake.

Each step keeps only the listings that passed the one before. Indexed is everything any source lists; the last step is our readiness grade A or B.

MCP servers

  1. Indexed61,862
    100.0% of indexed
  2. Has a remote endpoint61,239
    99.0% of indexed · 99.0% of previous
  3. Reachable53,021
    85.7% of indexed · 86.6% of previous
  4. Full handshake (initialize + tools/list)35,921
    58.1% of indexed · 67.7% of previous
  5. Readiness A or B34,478
    55.7% of indexed · 96.0% of previous
Measured 2026-11-12 01:39 UTC from the latest probe of every listing. Method JSON
02 · Sources

41.8% of MCP servers are listed by more than one source.

Which directories and registries list what we index, and how many of their listings answered our last probe. A listing found by several sources is counted once in the index.

MCP servers

SourceListingsReachable
mcp-registry41,05290.8%
agenstry19,53979.3%
own /.well-known crawl17,66093.5%
smithery13,41282.6%
capkey9,48496.6%
pulsemcp5,33180.0%
chiark2,98668.7%
apify80290.1%
Listed by more than one source: 41.8%
Source membership as recorded by our crawlers; reachability from the latest probe. Method JSON
03 · Protocols

7.6% of reachable MCP servers report the latest revision, 2026-07-28.

The protocol revision each reachable listing reports: the MCP version returned by initialize, the A2A version declared in the Agent Card, and the x402 version seen in the 402 challenge.

MCP protocol version

RevisionShareCount%
2026-07-28
4,016 7.6%
2025-11-25
18,697 35.3%
2025-06-18
5,496 10.4%
2025-03-26
7,043 13.3%
2024-11-05
1,401 2.6%
Non-standard
20 0.0%
Not reported
16,348 30.8%
Reachable listings only, as of 2026-11-12 01:39 UTC. Method JSON
04 · Readiness

55.8% of MCP servers grade A or B; 14.6% grade F.

Readiness grades whether an agent could call the listing today: protocol checks, declarations, uptime and freshness. It is shown on each listing and does not change search order, which uses the quality score.

ProtocolGrade mixABCDF
MCP servers 782
1.3%
33,696
54.5%
2,381
3.8%
15,974
25.8%
9,029
14.6%
Grades stored by the daily readiness job. Method JSON
05 · Reliability

Median response for MCP servers is 326 ms; p95 is 1175 ms.

30-day uptime counts only listings with at least five probes; probes a host refused (HTTP 429, or a 403 we were not let through) are left out. Latency is the last successful probe from our server.

MCP servers

Median latency
326 ms
p95 latency
1,175 ms

30-day uptime (59,329 tracked)

UptimeShareCount%
100%
47,160 79.5%
95–99%
2,544 4.3%
80–94%
1,475 2.5%
50–79%
726 1.2%
<50%
7,424 12.5%
Probe history of the last 30 days, as of 2026-11-12 01:39 UTC. Method JSON
07 · Access

72.8% of reachable MCP servers do not say how to authenticate.

How a caller is expected to authenticate, as declared by the listing. MCP servers that accept x402 are counted as pay-per-call even when they also take a key.

MCP servers

Declared accessShareCount%
Open (declared)
185 0.3%
API key or token
14,047 26.5%
OAuth
137 0.3%
x402 pay-per-call
30 0.1%
Not declared
38,622 72.8%
Reachable listings, declarations only; we do not attempt to log in. Method JSON
08 · Safety

36 MCP servers are flagged by our static safety scan.

A static scan of MCP tool names and descriptions for prompt-injection and data-exfiltration patterns. Flagged servers stay listed with a warning; a clean result is not a security review.

Verdicts (61,239 scanned)

clean
61,203
suspicious
35
malicious
1

Rules that fired

tool poisoning coercion
19
prompt injection
7
data exfil coercion
6
bare ip url
2
install lure
1
pipe to shell
1
Scan stored with each MCP listing. Method JSON
09 · Topics

'payment' is 8.7% of reachable A2A agents but 1.7% of MCP servers.

Share of reachable listings whose name or description mentions each topic word. Darker cells mean a larger share; compare across a row to see where one protocol has supply and another has little.

TopicMCP servers
ai 24.3%
search 11.7%
data 9.6%
github 4.4%
payment 1.7%
code 3.2%
web 2.4%
email 2.6%
video 2.6%
llm 1.5%
crypto 1.2%
wallet 0.9%
image 1.9%
pdf 1.6%
analytics 1.6%
document 1.4%
security 1.1%
social 1.4%
news 1.1%
knowledge 1.1%

Largest gaps

  • payment: 8.7% of A2A agents, 1.7% of MCP servers
  • wallet: 5.1% of x402 APIs, 0.9% of MCP servers
Keyword match on name and description; one listing can count for several topics. Method JSON
10 · Growth

3,600 MCP servers were first seen in the last 7 days; 540 went down.

Listings by the day any source first showed them to us, over the last 7 days. Our crawlers started at different dates, so early spikes partly reflect when a source was added.

MCP servers

New in 7 days
3,600
Went down in 7 days
540
2026-11-06: 4322026-11-07: 4842026-11-08: 5282026-11-09: 8622026-11-10: 6592026-11-11: 6352026-11-12: 0
2026-11-062026-11-12
First-seen dates from listing sources. Method JSON
11 · Adoption

GitHub data collected so far for 16,071 of 18,491 reachable MCP servers that link a repository; figures fill in over the coming days.

Signals from outside our probes: GitHub stars of the source repository, npm downloads, and where the site ranks in the Tranco top-1M list. They describe attention, not quality, and do not affect search order.

GitHub stars, reachable MCP servers (16,071 with a repo)

StarsShareCount%
0
11,165 69.5%
1–9
4,071 25.3%
10–99
543 3.4%
100–999
166 1.0%
1k–9.9k
104 0.6%
10k+
22 0.1%

Most-starred repositories

com.browser-use/browser-use
browser-use/browser-use
★ 120,690
io.github.IncorporatedPartners/labelhead-artist-momentum
paperclipai/paperclip
★ 113,080
World Monitor
koala73/worldmonitor
★ 90,482
com.ainetcafe/ai-netcafe
ChatGPTNextWeb/NextChat
★ 88,945
Netdata
netdata/netdata
★ 81,367
Front-End Checklist
thedaviddias/Front-End-Checklist
★ 74,706
Context7
upstash/context7
★ 64,598
io.github.tldraw/tldraw
tldraw/tldraw
★ 51,462
Reactive Resume
reactive-resume/reactive-resume
★ 45,629
PostHog MCP Server
PostHog/posthog
★ 40,740

Site rank in Tranco top 1M (list of 2026-11-10)

ProtocolTop 10k10k–100k100k–1MRanked
MCP servers 3225881,638 4.8%

Ranked by the domain an agent calls (MCP endpoint, x402 URL, A2A interface). Domains behind more than 20 of our listings, hosting platforms and content sites such as youtube.com or github.com are left out.

Most-downloaded npm packages (30 days)

firecrawl-mcp630,896
@launchdarkly/mcp-server273,391
tavily-mcp85,055
@brightdata/mcp73,483
@postman/postman-mcp-server46,496
GitHub and npm refreshed weekly; Tranco list refreshed daily. Method JSON
Scope & limits

Methodology

Every number is a count over the agent-tools.cloud index at the time shown above, recomputed once a day after our probes run. It is not a census of the whole ecosystem: listings no public source has shown us are missing, and private deployments are invisible to us.

Funnel. Reachable means the last probe got a valid answer. Verified means a valid 402 challenge for x402, a full MCP initialize plus tools/list, or an A2A JSON-RPC reply on the declared interface.

Sources. Duplicates across sources are merged by endpoint; each source a listing came from is kept.

Protocols. Versions are what the listing reports or declares; we do not infer them from behaviour. A2A signatures are verified against the key the card points to on its own site.

Readiness. Criteria and weights are on the About page. Points earned from live probes decay after 30 days without a fresh one.

Reliability. One probe per listing per day from one server; latency includes our network path and is not a provider SLA.

x402. Price per API is the lowest amount its payment terms ask for one call, in USD (USDC read at par). The x402 version is counted only where our resource probe recorded it. Settlement network falls back to the chains a listing declares when the challenge does not name one.

Access. Declared by the listing or its card; not tested with credentials.

Safety. Pattern rules over tool metadata only; no code is executed.

Topics. Whole-word match of a fixed word list on name and description. Listings whose text names more than five topics (gateway catalogues) are left out: 1,202 x402 APIs, 170 MCP servers, 54 A2A agents.

Growth. Earliest first-seen date across a listing's sources.

Adoption. GitHub stars and last push from the GitHub REST API for the repository a listing links as its source; npm downloads over 30 days from the npm downloads API for packages published on npm; site rank from the Tranco research list (registrable domain). Several listings can share one repository. These are attention signals, not measures of quality, and are not used for ranking.

Reuse

Cite and download

Free to reuse under CC BY 4.0 with a link to this page. The full snapshot is at /stats.json; headline numbers for every day are at /stats/daily.csv.

agent-tools.cloud. "The state of agent tools." 2026-11-12. https://agent-tools.cloud/stats

@misc{agenttools_stats_20261112,
  author = {agent-tools.cloud},
  title  = {The state of agent tools},
  year   = {2026},
  note   = {Snapshot 2026-11-12 01:39 UTC},
  url    = {https://agent-tools.cloud/stats}
}