AI・テクノロジー
MCP Surpasses 97 Million Installs | The Dominant Standard for AI Agents
機械翻訳 / Machine-translated
A
AI Friends
2026年4月26日・22分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
"We want to use AI agents internally, but every system integration costs us development time all over again…" — Companies grappling with this problem have reason to celebrate.
On March 25, 2026, Anthropic's MCP (Model Context Protocol) surpassed 97 million installs.
With OpenAI, Google, and Microsoft all on board, it has become the AI industry's "de facto standard."
MCP (Model Context Protocol) is an open standard that connects AI agents to external tools, databases, and applications in a standardized way. Anthropic announced it in November 2024.
Think of it as the USB-C port of the AI industry.
Before USB-C arrived, smartphone charging cables varied by manufacturer, forcing travelers to carry multiple cables on every trip.
The easiest way to understand MCP is this: it created a universal connector between AI and applications that solves that same fragmentation problem.
Specifically, an AI agent (client) connects to relay points called MCP servers, enabling it to call file systems, databases, SaaS tools, and internal tools using a common protocol. What makes this groundbreaking is that it eliminates the need to write three separate codebases — one for ChatGPT, one for Claude, one for Gemini.
In traditional AI application development, there was a problem known as the "N×M problem."
Connecting N AI models to M tools required writing N×M connectors.
For example, linking ChatGPT, Claude, Gemini, and Llama (4 AI models) to Slack, Notion, GitHub, and Jira (4 tools) required 16 different integration codebases.
MCP transformed this into an "N+M" problem.
With AI models supporting MCP on one side and tools providing MCP servers on the other, just 4+4=8 implementations cover all 16 possible combinations.
Development effort is cut by more than half, dramatically lowering the cost of AI adoption for businesses.
The "97 million installs" figure refers to monthly SDK downloads.
Shortly after MCP's release (late 2024), monthly downloads stood at 100,000.
That number ballooned 970× to reach 97 million by March 2026, just 18 months later.
As a point of comparison, consider Kubernetes, which became the standard of the cloud industry.
After Google released Kubernetes in 2014, it took approximately four years to reach equivalent adoption density in enterprise environments.
MCP covered the same adoption curve in just 18 months.
To use an analogy: a small shared bus stop built in one town grew in 18 months into a universal ticket gate installed at every major station in a metropolitan area. An infrastructure standard spreading this fast is a rare phenomenon even in the history of IT.
Three strategic factors explain why MCP became a standard so rapidly.
Imagine Sato-san (a pseudonym), who handles all IT for a logistics company with 80 employees.
In the past, whenever he wanted to "aggregate Slack inquiries into Claude, summarize Notion meeting notes, and auto-update GitHub issues," he had to write custom scripts calling three separate APIs — maintenance alone consumed 10 hours a month.
With MCP, he just configures the official MCP servers for Slack, Notion, and GitHub.
Then he simply tells Claude: "Summarize last week's Slack inquiries, turn them into Notion meeting notes, and file a GitHub issue." Everything gets done.
Operational effort drops to one-tenth of what it was before.
MCP is not the only mechanism for connecting AI to external tools. Here's how it compares to the main alternatives.
What's striking is that MCP was the only contender capable of becoming an industry standard.
Function Calling is convenient but carries lock-in tendencies with separate specs per company. OpenAI Plugins failed because other companies wouldn't join a single-vendor standard. MCP succeeded because it "promised openness from the start," letting even competing companies adopt it with confidence.
"Building an AI product specialized for a single model" → Function Calling is sufficient.
"Spanning multiple AI models and multiple tools" → MCP is the only choice.
"Existing proprietary API integrations are already running" → Gradually migrating to MCP is the practical answer for 2026.
Multiple companies in Japan are now moving seriously into MCP adoption.
Japanese white-collar workers spend enormous amounts of time switching between multiple SaaS applications.
For a salesperson, cycling between Salesforce, kintone, Slack, and Outlook dozens of times a day is a typical reality.
If a single AI assistant can operate all tools via MCP, a reduction of 200–500 hours per person per year in working time becomes a realistic prospect.
A. Yes, it is completely free and open source.
Both the specification and the SDK are published under the MIT License, and commercial use is unrestricted.
Because it is managed by the Agentic AI Foundation under the Linux Foundation, there is no risk of it suddenly becoming unavailable due to any single company's decisions.
A. MCP servers can be hosted within your own environment, so there is no need to send data externally.
The design returns only the results of tool calls to the AI model, and adoption is progressing even in financially and medically regulated sectors with strict security requirements.
In Enterprise versions of ChatGPT and Claude, input data is not used for model training.
A. For personal use, the "user side" is simple.
All you need to do is add the MCP server URL to the Claude Desktop app's configuration file, and file system and web search tools become instantly available.
"Building your own MCP server" is aimed at engineers, but with 17,000+ servers — both official and unofficial — already published, most users will find that combining existing solutions is more than enough.
A. Copilot is a "product"; MCP is a "standard."
Microsoft Copilot is the name of the AI product integrated into Microsoft 365. Since Microsoft Build in August 2025, Copilot has also added MCP support.
In other words, the relationship is not "Copilot vs. MCP" — it's "MCP running inside Copilot."
A. As of April 2026, it is expanding rapidly.
Official MCP servers for major domestic SaaS products — including kintone, Cybozu, and freee — are being released in quick succession, and large firms such as Hakuhodo DY have begun offering implementation support services.
Within a year, it is expected that most tools used by Japanese companies will have MCP support.
Just as the arrival of the USB-C port instantly made the gadget world more convenient, MCP is poised to bring the same revolution to the world of AI and business systems.
"How quickly can we connect our company's AI to our internal tools?" will be the dividing line in the productivity competition of the next two to three years.
The step you and your team take this week may create a decisive difference in your operational speed one year from now.
This article is a cross-post from AI Friends.
