MCP Adoption Statistics 2026: Downloads, Servers and A2A

Most MCP statistics I find are one number in one article: a server count from a directory, a download figure from a launch post, a vendor list from a conference keynote. Each was true on the day it was written. None of them tells you whether adoption is speeding up or flattening out.

This page measures the trend. A script reads the official MCP Registry, the npm and PyPI download logs and GitHub every day, keeps every reading, and rebuilds the numbers below. The vendor table is checked by hand against each company's own announcement. A2A, the agent-to-agent protocol Google introduced in April 2025, sits alongside so you can compare the two on the same measures.

It is a census, not a directory. If you want to browse or search servers, Agenstry, the official registry and Smithery do that well. If you want the concepts first, AI agent vs MCP server explains what a server is, and MCP vs A2A vs ACP explains how the protocols differ.

MCP vs A2A at a glance

The same measures for both protocols, read on the same day. Where a measure does not exist for A2A, such as an official server registry, the row says so.

All figures read on 4 October 2026 unless the row gives another date. Downloads are raw registry counts, which include CI systems and dependency installs, so they measure how widely the SDKs are built into software, not how many people use MCP. GitHub topic counts cover public repos that chose to tag themselves.

SDK downloads over time

Weekly downloads of the official SDKs are the longest clean series available for either protocol: npm keeps daily counts back to the TypeScript SDK's first release in November 2024, and the PyPI logs go back as far. The chart covers the last 52 weeks.

Two things stand out in the series. In October 2026, MCP's weekly downloads ran about 20 times A2A's on both registries. A2A is growing faster in percentage terms, from a far smaller base, which is what you would expect from a protocol adopted mainly inside enterprise agent platforms rather than by individual developers wiring tools to a model. The year-on-year figures for both are in the table above.

The official MCP Registry

The MCP Registry launched in preview on 8 September 2025 as the protocol's own catalogue, with a public API. Publishers register a server under a namespace they can prove they own, either a domain or a GitHub account. That makes it the cleanest growth signal for MCP: each entry is a deliberate act by a publisher, not a crawler's guess.

Read from the public /v0/servers API of the official registry on 4 October 2026. A server can list both a remote endpoint and packages, so shares overlap. "First published" uses the earliest publish date of any version of the server, as the registry reports it.

The shape is steep. The registry took until January 2026 to pass 1,500 servers, then added more in September 2026 alone (11,902) than in its first nine months combined (10,298). Most listed servers are remote: endpoints a client connects to over HTTP, with no package to install. That fits how MCP is now used, with hosted products exposing their own server rather than developers running a local process.

Which vendors support MCP and A2A

One row per vendor, each "Yes" linked to the vendor's own announcement, docs, release notes or press release, with the earliest date I could confirm. I record what was announced, not what any one product does today, and I list A2A launch partners separately from vendors that shipped A2A in a product.

Each "Yes" links to the vendor's own announcement, docs or changelog, checked on 4 October 2026. "Launch partner" means the vendor was named among A2A's launch partners, which is a statement of intent rather than a shipped product. "Not confirmed" means I found no primary source; it is not a claim that support is absent.

The pattern is clear even with the caveats. MCP support is close to universal among large AI and software vendors, and it came early: most of the dates fall in the first half of 2025. A2A support is concentrated in enterprise agent platforms, where agents built by different vendors need to hand work to each other. Several model providers, including Anthropic, OpenAI, Mistral and xAI, have no A2A announcement I could find.

Change log and timeline

Protocol governance events, each with its primary source, followed by the movements the daily data picks up: registry milestones and GitHub topic milestones.

1. 4 October 2026: Census launched. 39,238 servers in the official registry, 32,944 GitHub repos tagged mcp-server.
2. 9 December 2025: Anthropic donates MCP to the new Agentic AI Foundation under the Linux Foundation; platinum members include AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI (source)
3. 8 September 2025: MCP Registry launches in preview (open catalog and API for MCP servers) (source)
4. 29 August 2025: IBM Research's Agent Communication Protocol (ACP) merges into A2A under LF AI & Data (source)
5. 1 August 2025: A2A v0.3 (gRPC, signed agent cards); Google says 150+ organizations support A2A (source)
6. 23 June 2025: Linux Foundation launches the Agent2Agent Protocol Project; Google contributes A2A, with supporting quotes from AWS, Cisco, Microsoft, Salesforce, SAP and ServiceNow (source)
7. 9 April 2025: Google announces the Agent2Agent protocol (A2A) with 50+ partners (source)
8. 25 November 2024: Anthropic introduces and open-sources the Model Context Protocol (source)

What the numbers do not show

Downloads are not users. npm and PyPI count every install, including CI runs, container builds and packages that depend on the SDK. The weekly figures say how deeply MCP is built into software, and they are good for comparing trends. They are not a count of people.

Registry size is not the size of the ecosystem. Many servers are never published to any registry, and many listed servers are experiments. Crawlers that index every public endpoint they find report far more than the official registry: on 4 October 2026 Agenstry showed 204,020 MCP servers indexed and 12,579 responding to its live checks, and 7,229 A2A agents with a valid agent card. Those counts answer a different question, how much is out there, while the registry answers how fast publishers are choosing to list.

A vendor announcement is not a product you can use today. Some A2A rows record a preview or a "coming soon". The table dates the first announcement because that is the verifiable event.

I run Gravity, an agent platform in private alpha, and agents on it call tools, so MCP's adoption matters to me in a practical way. I have no stake in which protocol wins, and the page reports both on the same footing. For what agent platforms charge, see the AI agent pricing index and the cheapest AI agent platforms.

Methodology

• Registry size: every server in the official registry's public /v0/servers API, latest version of each, read daily and paged to the end. Growth uses each server's first-publish date: for servers listed on 4 October 2026, the earliest publish date of any version (from a full crawl of all 132,432 published versions); for servers that appear later, the day this tracker first sees them. Servers removed from the registry before 4 October 2026 are not counted, so early months may read slightly low.
• Smithery: the totalCount its public registry API returns, read daily.
• SDK downloads: daily counts from the npm downloads API for @modelcontextprotocol/sdk and @a2a-js/sdk, and from ClickPy, ClickHouse's public copy of the PyPI download logs, for mcp, fastmcp and a2a-sdk. The PyPI name mcp belonged to an unrelated package before the SDK took it over, so the series starts in November 2024. Weekly figures are the last seven complete days; year-on-year compares with the seven days ending 364 days earlier.
• GitHub: repository search counts for the topics mcp-server, model-context-protocol, mcp, a2a-protocol and a2a, read daily. The monthly history in the CSV counts repos that exist today by the month they were created, so deleted repos and topics added later blur it slightly. Stars on the spec and SDK repos come from the GitHub GraphQL API.
• Vendor support: checked by hand on 4 October 2026 against each vendor's own blog, docs, changelog or press release, or the protocol's official announcement. A row says "Not confirmed" when I found no primary source. Corrections are welcome at support@gravity.fast with the source link.

The refresh runs daily. Sibling dataset: the open-source AI agent framework leaderboard uses the same collector for 43 agent frameworks. For broader numbers on agents, see AI agent statistics for 2026.

FAQ

Sources