MCP Crosses 97 Million Installs: How One Protocol Became the Universal AI Connector

When Anthropic open-sourced the Model Context Protocol in late 2024, the skeptics had a fair point: why would the industry coalesce around one company's connector standard when every platform was building its own tool-calling interface? Fourteen months later, the numbers answer that question emphatically.

MCP has crossed 97 million installs as of March 2026. That figure, drawn from npm registry data and package manager telemetry, represents the combined installations of the TypeScript, Python, Java, and Kotlin SDKs across every environment where MCP is used — Claude Desktop, Claude Code, VS Code Copilot, Cursor, Windsurf, Cline, and dozens of smaller hosts.

The Adoption Timeline

The growth curve tells the story of a protocol that found product-market fit across an entire industry:

Three inflection points stand out:

March 2025: OpenAI adopts MCP. When the Agents SDK shipped with MCP client support, it validated the protocol for the entire ecosystem. Developers who were building OpenAI-specific tool integrations suddenly had a reason to standardize on MCP instead.

September 2025: IDE explosion. VS Code Copilot, Cursor, Windsurf, and JetBrains AI Assistant all shipped MCP support within the same quarter. This moved MCP from a niche Anthropic feature to something every developer encountered daily.

January 2026: Streamable HTTP transport. The release of remote MCP servers via Streamable HTTP unlocked enterprise use cases. Teams could deploy shared MCP servers instead of requiring every developer to run local instances. Corporate adoption accelerated dramatically.

Who Is Using MCP

The install distribution reveals a protocol that has penetrated every layer of the AI development stack:

IDE extensions account for the largest share. This makes sense — MCP's original value proposition was connecting AI coding assistants to development tools (databases, APIs, documentation). That use case has not changed, but it has scaled to millions of developers.

The enterprise segment, while smallest by install count, is growing fastest. Companies are deploying internal MCP servers that give their AI assistants access to proprietary data — CRM systems, analytics dashboards, internal wikis — through a standardized, authenticated protocol.

The Remote Server Revolution

The most significant recent development is the shift from local to remote MCP servers. For the first fifteen months of MCP's existence, servers ran as local processes communicating over stdio. This worked for personal tools but created friction for teams and enterprises.

The Streamable HTTP transport, finalized in late 2025 and widely adopted in early 2026, changed the calculus:

Remote servers run as standard HTTP services — deployable on Cloudflare Workers, AWS Lambda, Fly.io, or any container platform. They support bearer token and OAuth authentication, integrate with existing monitoring infrastructure, and update centrally without touching client machines.

Cloudflare has emerged as a particularly popular host. Their Workers platform, with edge deployment, D1 database bindings, and zero cold start overhead, maps almost perfectly to the MCP server use case. Cloudflare published official MCP deployment templates in February 2026, and the community response has been substantial.

What the Competitors Did

The most telling sign of MCP's dominance is who capitulated:

OpenAI — Initially built their own tool-calling format for GPT function calling. In March 2025, they shipped MCP client support in the Agents SDK. By September 2025, their recommended approach for tool integration was MCP-first.

Google — Shipped MCP support in Gemini's agent framework (Project Mariner predecessor) in Q3 2025. Google's DeepMind research team has contributed to the MCP specification, particularly around streaming and multi-modal content types.

Microsoft — VS Code Copilot gained MCP support in August 2025. GitHub Copilot Workspace followed in November. Microsoft is now one of the largest contributors to the MCP specification working group.

Vercel — Their AI SDK added MCP client support, making it trivial for Next.js applications to consume MCP tools. This brought MCP into the web application layer beyond just IDE usage.

The Protocol War That Wasn't

Before MCP, there were at least four competing approaches to connecting AI to tools:

1. OpenAI Function Calling — JSON Schema-based, model-specific
2. LangChain Tool Definitions — Framework-specific, Python-centric
3. Semantic Kernel Plugins — Microsoft's .NET-centric approach
4. Custom REST wrappers — Everyone's first attempt

MCP won because it solved the N-by-M problem. Without a standard, every AI host (N) had to build custom integrations for every tool (M). With MCP, tools are written once and work everywhere. The value of the protocol increases with every new host that supports it — classic network effects.

The 97 million install number is not just adoption. It is a moat. Every developer who writes an MCP server is creating a tool that works with every AI assistant on the market. That creates a gravity well that makes competing protocols increasingly pointless.

What Comes Next

The MCP specification is actively evolving. The 2026 roadmap includes several changes that will accelerate enterprise adoption:

Server discovery via .well-known/mcp endpoints will let clients auto-discover what tools a server offers without connecting to it. This enables marketplace-style discovery of MCP servers.

Stateless protocol improvements will decouple sessions from transports, making it trivial to run MCP servers behind load balancers and auto-scaling groups.

Multi-modal content types will expand tool responses beyond text to include images, structured data, and interactive elements.

The question is no longer whether MCP will be the standard for AI tool integration. It already is. The question is how far the standard extends — and whether 97 million installs is the beginning or the midpoint of the growth curve.

For a hands-on guide to building on this protocol, see our tutorial on deploying a remote MCP server with Streamable HTTP on Cloudflare Workers.