· via dev.to (home feed)
MCP SDK downloads triple to 97 million a month, dev.to analysis finds
A dev.to survey of the MCP ecosystem reports 97 million monthly SDK downloads, triple the level of six months ago, and more than 13,000 servers — but discovery remains the weak spot.

What the numbers show
A dev.to post published on 3 October 2026 rounds up growth figures for the Model Context Protocol ecosystem, and the headline number is big: roughly 97 million monthly downloads of MCP SDKs, which the author says is about triple the level of six months earlier.
The author, describing an afternoon spent digging through npm and GitHub, reports the following:
- More than 13,000 MCP servers listed across npm and GitHub as of May 2026.
- Around 97 million monthly SDK downloads, roughly three times the figure from six months prior.
- A 400% year-over-year increase in new server registrations.
- Anthropic's official filesystem server alone drawing about 48,500 downloads per month.
These are self-reported figures from a single informal survey rather than an audited count, and the post offers little detail on methodology, so they are best read as directional.
From protocol to plumbing
MCP is the open protocol that standardizes how AI models and agent applications connect to external systems such as databases, APIs and file systems, rather than each integration being built bespoke. The dev.to author's argument is that the protocol has crossed a threshold: it is no longer a specification under discussion but the default wiring for tool-using agents. The download curve is the evidence offered, since SDK demand tripling in half a year points to developers wiring MCP into shipping applications rather than merely evaluating it.
Discovery is the bottleneck
The post's sharpest observation is not about growth but about friction. With more than 13,000 servers in circulation, the hard part is no longer connecting tools to models, according to the author — it is finding the right server in the first place. The current experience is described as guesswork searches on npm or digging through folders on GitHub, with no straightforward path from needing a Postgres connector to a known-good package.
A registry ships alongside the analysis
The post also doubles as a launch announcement. The author has released mcp-hub, a command-line tool for searching and installing MCP servers: install it globally through npm, search for a term like "database", then install a package such as the official Postgres server. The registry launches with five official servers, backed by real packages the author says were verified on npm. Planned next steps include private registries for enterprise teams, community submissions via GitHub issues, and CI/CD integration for auto-publishing servers.
That framing matters for how the numbers are read. The growth survey and the product it accompanies come from the same post, so the statistics have not been checked by anyone without a stake in the ecosystem's momentum.
Why it matters
If the figures are even approximately right, they mark a real shift in agent tooling. Tripling SDK downloads in six months is the kind of curve that precedes a protocol becoming infrastructure — the layer developers assume rather than actively choose. For anyone building agents, that means tool integrations are increasingly written once against MCP and reused across models and clients.
The growth also relocates the problem. Once a protocol has thousands of implementations, the scarce goods become discovery, quality control and trust: which server is maintained, which one should be handed credentials, and which satisfies enterprise requirements. Curated registries with verification, private distribution and automated publishing — precisely the roadmap the dev.to author sketches — look like the next part of the ecosystem to consolidate.
The caveat stands. This is a single, informal count published alongside a self-promotional tool launch, so the direction of travel is more trustworthy than the precise digits. Independent confirmation from npm or Anthropic would turn an interesting anecdote into a hard data point.
- #model-context-protocol
- #mcp
- #ai-agents
- #npm
- #developer-tools