The State of MCP Servers 2026

The Model Context Protocol registry now lists 13,108 servers — but the ecosystem is a long tail with almost no middle. Analyzing the 2,500 most-starred servers, we found it is overwhelmingly local-first (83% run on your machine, not in the cloud), radically decentralized (2,102 publishers, 91% shipping a single server), and JavaScript-first (npm outnumbers Python 2-to-1).

Model Context Protocol (MCP) went from a late-2024 spec to the default way agents talk to tools. Eighteen months in, the official registry has crossed 13,000 servers. To understand what that ecosystem actually looks like — not the marketing, the data — we pulled the registry and computed the numbers on its most-visible 2,500 servers. Here is what the ecosystem is made of. You can browse the underlying catalog any time at our MCP directory.

Key findings

• 83% of MCP servers run locally. 2,074 of 2,500 use the stdio transport (on-machine); only 14.8% are remote HTTP/SSE servers.
• The registry lists 13,108 servers, but the curve collapses fast. Even inside the top 2,500 by stars, the 500th-ranked server has just 59 stars and the 2,500th has 2.
• The median MCP server has 10 GitHub stars. 84% have fewer than 100; only 4.1% clear 1,000 stars and just 1.0% clear 10,000.
• The ecosystem is radically decentralized. 2,102 distinct publishers account for 2,500 servers, and 90.6% of publishers ship exactly one server. The top 10 publishers combined are only 5.7% of the total.
• npm is the dominant distribution channel. 52.2% of servers ship an npm package versus 23.2% PyPI — MCP is a JavaScript-first ecosystem, with Docker/OCI (8.7%) a distant third.
• The most-starred server is Netdata (79,443 stars) — though most top entries are large host products that happen to contain an MCP server, not standalone MCP projects.
• It's a live ecosystem, not a graveyard. 66% of these servers had a code push in the last 30 days; 96% within six months; none older than a year.
• A single MCP bundle format (mcpb) already appears on 4% of servers — an early signal that one-click desktop install is taking hold.

Chart to accompany this section: a summary "stat card" row (13,108 total · 83% local · 91% single-server publishers · 10 median stars).

Finding 1: MCP is still a local-first protocol

The single clearest signal in the data: 83.0% of servers use stdio, the transport that runs a server as a local subprocess on the same machine as the agent. Only 14.8% are remote (14.0% HTTP, 0.8% SSE), with 2.3% undeclared.

Caption — chart: horizontal stacked bar, stdio (local) vs HTTP/SSE (remote) vs unknown, 100% width.

The picture people have of MCP — remote, hosted, OAuth-gated cloud services — is not what the ecosystem is built from today. The overwhelming majority of MCP servers are local binaries you install and the agent launches. Remote-first MCP is the emerging frontier, not the baseline. It is also why the platform you run agents on matters: whether a host can launch a local stdio server or only reach remote HTTP endpoints decides which half of this ecosystem you can actually use (see our platform directory).

Finding 2: 13,108 servers, but a very long tail

The registry advertises 13,108 servers. We analyzed the top 2,500 by GitHub stars — and the drop-off inside even that popular slice is severe. Ranked by stars, the list falls from 79,443 at #1 to 59 stars at rank 500, 16 at rank 1,000, and 2 stars at rank 2,500. Everything below our cut-off — over 10,000 servers — has essentially zero stars.

Caption — chart: log-scale line, stars by rank (1 to 2,500), annotated at ranks 1 / 500 / 1,000 / 2,500.

This is a classic power-law ecosystem: a handful of prominent servers, a thin shoulder of moderately-known tools, and a very long tail of experimental, single-author, or abandoned projects.

Finding 3: The median server has 10 stars

Aggregate popularity confirms the long tail. Across the 2,500 servers with star data, the median is 10 stars while the mean is 496 — the gap between them is the fingerprint of a skewed distribution dragged up by a few giants.

Caption — chart: histogram of star buckets (0-9, 10-99, 100-999, 1k-9.9k, 10k+) with the median marked.

For anyone choosing an MCP server, the practical takeaway is that "it's in the registry" tells you almost nothing about maturity — the vast majority are early-stage. Our best-of shortlists exist for exactly this reason.

Finding 4: No dominant vendor — 91% of publishers ship one server

MCP has no equivalent of a platform gatekeeper. 2,102 distinct publishers produced these 2,500 servers, and 1,905 of them (90.6%) published exactly one. Concentration is remarkably low.

Caption — chart: Pareto/lollipop of the top 15 publishers with a "top-N cumulative share" overlay.

Even the single largest "publisher" — ai.smithery, at 38 servers — is an aggregator/hosting platform, not a product company. The most-prolific vendors trail well behind. This is an ecosystem built by thousands of individual authors, not a few large suppliers.

Finding 5: JavaScript won the distribution layer

Where do MCP servers actually ship? By packaged distribution channel:

Caption — chart: bar chart, servers by distribution channel (npm, PyPI, Docker, mcpb, NuGet).

npm outnumbers Python more than two-to-one. Despite MCP's reference implementations spanning both TypeScript and Python, the community has concentrated on the JavaScript toolchain (npx-launchable servers) as the path of least resistance. The appearance of mcpb — the MCP desktop-extension bundle format — on 99 servers is a small but notable signal that one-click install is arriving.

Finding 6: The ecosystem is fresh, not abandoned

A common worry with fast-growing registries is rot. Not here — at least not at the top. Measuring the last code push per server:

Caption — chart: cumulative area, % of servers by "days since last push" bucket.

The popular slice of MCP is under active development. (This measures the top 2,500 by stars, which skews toward maintained projects — see Methodology — but among the servers users are most likely to reach for, staleness is rare.)

Finding 7: Category data is coarse — but specialization is emerging

The registry's category signal is blunt: 82.7% of servers fall into "Development." That partly reflects a coarse classifier, and partly reflects the genuine reality that most early MCP servers are developer tools. More interesting is what rises above the noise as distinct specializations.

Caption — chart: treemap of categories, "Development" muted so the specialized verticals stand out. (14 categories total.)

Search & Web (4.3%), Databases (1.8%), AI & ML (1.7%), and Communication (1.0%) are the four clearest non-developer verticals forming today. For a closer read on where the ecosystem's gravity sits, see our earlier analysis, what the most-starred MCP servers tell us.

Methodology

Dataset. Server metadata from the official MCP registry (registry.modelcontextprotocol.io/v0/servers), enriched with GitHub star counts and repository push dates. Snapshot generated 2026-07-01.

Sample. The registry lists 13,108 servers; this analysis covers the 2,500 most-starred (the dataset is star-sorted descending, which we verified). Every statistic here describes that popular head unless noted. Because the sample is star-ranked, findings about freshness and popularity are biased toward maintained, visible projects — the true long tail (the ~10,600 servers below our cut-off) is almost entirely near-zero-star and would push medians lower and staleness higher.

All figures were computed programmatically from the raw JSON (a Node script over the servers[] array), not estimated. Percentages are shares of the 2,500-server sample.

Honest limitations.

• Transport is self-declared in the registry; 2.3% of servers declare none, and a remote server can also ship a local package, so "local vs remote" reflects the primary declared transport.
• Auth is not reliably recorded. The registry marks 84.4% of servers none and 15.6% unknown — it does not capture OAuth vs API-key requirements, so we deliberately publish no "X% require OAuth" claim. Most local servers take credentials via environment variables the registry never sees.
• isOfficial is unusable (only one server is flagged) and testStatus is untested for all servers — The Agent Examiner has not yet independently exercised them.
• Stars measure the host repository, which for large products (Netdata, tldraw, Metabase) reflects the whole product, not the MCP server alone — so top-server rankings overstate MCP-specific popularity.
• Categories come from an automated classifier and should be read as directional.

Cite this report

This report was compiled by The Agent Examiner, the MCP-server research desk. We maintain the underlying dataset to help teams evaluate agent tooling, and we publish the aggregate numbers openly because the ecosystem is moving faster than anyone's mental model of it. The script and field definitions are reproducible; if you cite these figures, a link back to this report is all we ask. We plan to refresh it quarterly as the registry grows.

Suggested citation: The Agent Examiner, "The State of MCP Servers 2026," /news/state-of-mcp-servers-2026 (snapshot 2026-07-01).