MCP crosses 97M installs — the agent standard is locking in

97M

Anthropic's Model Context Protocol (MCP) passed 97 million cumulative installs by end of March — sixteen months after launch. What began as a Claude-only tool-use spec is now the default agent integration layer, with OpenAI, Google, and Microsoft all shipping compatibility.

This is rare. A protocol designed by one company has become the industry's plumbing.

Quick refresher on MCP

Model Context Protocol was released by Anthropic in November 2024. The idea is direct: stop writing N×M integrations between LLMs and external tools. Define one wire protocol — JSON-RPC 2.0 over stdio, SSE, or Streamable HTTP — and let hosts (LLM apps), clients (in-host tool callers), and servers (the file/DB/API exposers) speak it.

It took off because the alternative was painful. Every new LLM app that wanted to read GitHub, query Postgres, or post to Slack had to build a fresh integration. MCP turned that N×M problem into N+M.

[IMG#1]

Where the 97M came from

Category	Estimated installs	Share
Dev tools (GitHub, Filesystem, Shell)	38M	39%
Databases (Postgres, SQLite, MongoDB)	17M	18%
SaaS (Slack, Notion, Linear, Jira)	14M	14%
Browser/Scraping	11M	11%
Cloud infra (AWS, GCP, Azure)	7M	7%
Other (hobby, experimental)	10M	10%

Two takeaways. Dev tools dominate at 39%, driven by the Claude Code and Cursor IDE-integration boom. Databases plus SaaS together are roughly a third — meaning MCP is no longer a Claude demo, it's running in enterprise back offices.

The compatibility list

OpenAI, in April last year, added an MCP compatibility layer to its Tools API. Sam Altman has publicly said "We chose MCP because it works."

Google made MCP a first-class citizen in Gemini 2.5 in September 2025. Sundar Pichai called it "the lingua franca of the agent ecosystem."

Microsoft built MCP adapters into Copilot Studio and GitHub Copilot Workspace, so any MCP server can be attached without custom integration.

[IMG#2]

How standards travel from one company to all

Few protocols make this trip. HTTP did, then handed governance to W3C. gRPC did, then went to CNCF. GraphQL did, then went to the Linux Foundation. Common pattern: company ships → 1-2 years of adoption → governance handoff to a neutral body.

MCP hasn't taken that final step. Anthropic runs the GitHub org but there's no committee, no charter. OpenAI and Google support the protocol but have no governance voice. That's the biggest political risk over the next 12 months.

If Anthropic moves MCP under the Linux Foundation or OpenJS, the standard becomes durable. If it doesn't, Google's A2A protocol gradually carves out an alternative track.

Counter-moves

Google A2A — agent-to-agent communication. MCP is tool-to-agent; A2A is agent-to-agent. They overlap in some areas already.

OpenAI Function Calling — keeps its native standard but layers MCP adapters over it, attempting to own both lock-in and compatibility.

LangChain Agent Protocol — open-source consortium effort, but adoption is roughly 1/10 of MCP's.

Stakes

• Wins: Anthropic — owns a standard larger than its model business.
• Wins: Developers — one integration runs on any LLM; integration cost down to one-third.
• Loses: Closed-standard camps — would-be lock-in advantages erode.
• Watching: Governance — committee handoff makes it permanent; no handoff invites fragmentation.

Skeptical view

Simon Willison: "MCP works, but its security model is weak — arbitrary servers attaching to arbitrary LLMs creates fuzzy authorization boundaries. Enterprise needs an OAuth-style auth layer first."

Drew Breunig: "97M installs blends hello-world experiments with real production. Active-server count is the more meaningful metric."

What changes for you

For builders — start new integrations as MCP servers. Write once, run on Claude, GPT, Gemini. Over 100 official servers are public.

For founders — the "agent integration" category should assume MCP as default. Don't ship a competing standard.

For investors — Anthropic's model business gets headlines, but standard ownership is the quieter durable asset. Watch for licensing-policy shifts.

For end users — the MCP marketplace in Claude Desktop lets you wire calendar, mail, and notes into one chat surface.

3-Line Summary

• MCP crossed 97M cumulative installs by end of March; OpenAI, Google, and Microsoft all ship compatibility.
• A rare case of one-company protocol becoming an industry default — governance handoff is the open question.
• Integration cost drops to roughly one-third; Anthropic owns a strategic asset bigger than its model lead.

References

• Anthropic — MCP official site
• Anthropic Newsroom — installs milestone
• GitHub — MCP official servers
• Simon Willison — MCP security analysis
• The Verge — MCP coverage