Research · data from 30 October 2026
The MCP ecosystem by the numbers, October 2026
Who publishes MCP servers, what they run on, how many people use them. Five publishers account for 15% of the graded registry; 61% of servers with a public repository have zero GitHub stars; 94% of publishers have shipped exactly one server.
Key numbers
servers graded in the official registry, from 6,333 graded in September. 43,717 grade A to C.
of the graded registry comes from five publishers, each with more than 20 servers in a 3,000-server sample. Scaled: about 6,700 servers.
of servers with a public repository have zero GitHub stars. 4% have a hundred or more.
of publishers in the sample have shipped exactly one server (2,139 of 2,267 namespaces).
The registry is young, lopsided and mostly unread. Most of it was published in the last few months, a handful of accounts publish in bulk, and almost nobody has claimed, reviewed or starred the typical server. That is not a criticism of the servers; it is the state of a directory that grew from 1,200 to 48,000 entries in ten months.
Who publishes
2,267 distinct namespaces appear in the 3,000-server sample. 2,139 of them appear once. At the other end, five namespaces contribute 458 servers between them:
io.github.sadri-dridi167 · 6%io.github.pipeworx-io125 · 4%io.github.mcp-dir73 · 2%io.github.mrfentmen70 · 2%io.github.CSOAI-ORG23 · 0.8%
These are generated catalogues: one wraps hundreds of public APIs as read-only remote servers, another publishes a local package per data source, a third republishes the same template with a different name. They are not malicious and most grade A or B. They are why counting servers overstates the ecosystem and why this site ranks at most two servers per publisher on any list.
39% of servers come from verified reverse-DNS namespaces (a company domain proven by DNS), 61% from GitHub namespaces. The security edition compares how the two groups grade.
How servers are shipped
- Remote only (a URL you connect to)1,632 · 54%
- npm package624 · 21%
- Package and remote305 · 10%
- PyPI package282 · 9%
- MCPB bundle77 · 3%
- Docker image52 · 2%
- NuGet, Cargo, unknown28 · 0.9%
63% of graded servers have a remote endpoint and 35% are packages only. In the registry's first months the split was the other way; the shift to hosted servers is the biggest structural change of 2026 and the reason sign-in, not code, is now the main risk. 63% of sampled servers are callable through the gateway without installing anything.
Who uses them
| GitHub stars | Servers | Share |
|---|---|---|
| 0 | 1,127 | 61% |
| 1 to 9 | 501 | 27% |
| 10 to 99 | 144 | 8% |
| 100 to 999 | 61 | 3% |
| 1,000 or more | 16 | 1% |
The median server has zero stars. Fewer than one in twenty has a hundred, and 16 in 1,849 have a thousand. Weekly package downloads tell the same story where they exist: a median of 55 a week. Reviews are rarer still: 1 of 3,000 sampled servers has a user review and 1 has been claimed by its publisher; across the whole site there are 5 reviews. Adoption of MCP as a protocol is large (the official SDKs pass hundreds of millions of downloads a month); adoption of any individual server is tiny, and concentrated in a few dozen names.
Maintenance is fresher than the star counts suggest. The median server's last commit was 36 days ago, 31% were pushed in the last month, and only 7% have gone six months without a commit. A registry this new has not had time to be abandoned.
Licences
- MIT1,357 · 45%
- No licence detected1,289 · 43%
- Apache-2.0169 · 6%
- Non-standard (NOASSERTION)119 · 4%
- AGPL-3.0 / GPL-3.039 · 1%
- ISC, BSD and other12 · 0.4%
43% of servers have no detectable licence, almost all of them remote-only listings without a repository, which is a legal question for any company whose agents call them. Where there is a licence it is overwhelmingly MIT (79% of licensed servers).
What they are for
| Category | Share of registry | Remote | Tools per server |
|---|---|---|---|
| Everything else | 55% | 66% | 16.4 |
| AI and agents | 9% | 49% | 10.7 |
| Web search | 8% | 67% | 8.5 |
| E-commerce | 4% | 70% | 12.8 |
| Finance | 4% | 63% | 11.9 |
| Media | 3% | 68% | 10.0 |
| Productivity | 3% | 56% | 13.8 |
| Developer tools | 2% | 46% | 17.1 |
| Databases | 2% | 38% | 15.3 |
| 2% | 68% | 13.1 | |
| Sales and outbound | 2% | 82% | 25.8 |
| Scraping and browsing | 2% | 41% | 15.2 |
| SEO | 2% | 80% | 9.0 |
| Security | 2% | 60% | 10.8 |
| Maps and travel | 1% | 71% | 7.8 |
More than half the registry does not fit a product category: single-purpose wrappers, demos, personal tools, and the bulk catalogues above. Of the named categories, AI and agents, web search and e-commerce are the largest. Sales servers carry the most tools each (about 26); SEO and maps servers the fewest, under ten.
Are they up
of remote servers answered every one of their last 28 checks.
answered fewer than 95% of checks; 1% fewer than half.
median response to initialize. 90th percentile 1,079 ms, 99th 2,865 ms.
Remote MCP servers are more reliable than the stars would predict: nine in ten answer every probe. The slow tail is real, though. One server in ten takes over a second to say hello, which an agent pays on every tool call.
In context
The official registry opened in preview on 8 September 2025 and has not announced general availability. Independent censuses put it at roughly 43,000 to 47,000 active names in October 2026, with around 12,000 new servers published in September 2026 alone; the counts here are consistent with those. The protocol's own SDKs report close to half a billion downloads a month. Against that, the picture of individual servers above, zero stars, one publisher, no reviews, is the normal shape of a very new package ecosystem rather than a sign of failure: npm's long tail looked the same.
Comparable directories report different totals because they count different things: scraped GitHub repositories, mirrored marketplaces, or skills alongside servers. This page counts registry entries with a safety grade, once each.
Method and limits
- Population. 48,447 graded servers from the official MCP registry on 30 October 2026, of which 43,717 grade A to C and appear in the sitemaps. Grade counts, the gateway-ready count (32,013) and the skills count (4,675) are exact.
- Sample. 3,000 of the 43,717 drawn uniformly at random and read through the public listing API; proportions are within ±1.8 points of the population 95 times in 100. Figures "scaled to the catalogue" multiply a sample share by 43,717 and are rounded to the nearest hundred.
- Stars, downloads, last push. From GitHub and the npm/PyPI download APIs, refreshed on each scan; known for 1,849 of 3,000 (the ones with a readable repository or package). Servers with no repository have no star count and are excluded from star figures, which therefore flatter the catalogue.
- Uptime and latency. The gateway Worker probes every remote endpoint about four times a day; the figures are the last 28 checks per server.
- Publishers. A publisher is a registry namespace (
io.github.nameor a verified reverse-DNS domain). Counts of "single-server publishers" are within the sample; the true number of single-server namespaces in the whole catalogue is higher.
Use this data
Quote any number on this page with a link back to it. Suggested citation: MCP Market, "The MCP ecosystem by the numbers, October 2026", mcp.market/research/mcp-ecosystem-by-the-numbers-october-2026.
Journalists and researchers: the row-level sample behind this edition (CSV, no personal data) and cuts by category, publisher or transport are available on request from hello@mcp.market, usually the same day. We will also re-run any number on the current catalogue.
Earlier editions: all research.