MCP Hits 97 Million Installs: How an Open Protocol Became AI Agent Infrastructure
Model Context Protocol (MCP) surpassed 97 million installs in Q1 2026, evolving from Anthropic's experimental spec into a cross-industry standard supported by every major AI vendor. This piece reviews MCP's architecture, transport-layer evolution, ecosystem data, and competitive landscape.
In March 2026, the Model Context Protocol (MCP) — first proposed by Anthropic in late 2024 — officially crossed 97 million global installs. In less than 18 months, an open protocol went from a developer curiosity to an infrastructure layer natively supported by OpenAI, Google, Microsoft, Meta, and virtually every major AI vendor.
One Protocol, One Old Problem
Large language models are powerful, but they are trapped inside a chat window. Want one to query a database, read a file, or send a Slack message? Each integration requires bespoke API glue code. MCP exists to eliminate that inefficiency.
At its core, MCP is a client-server protocol that defines a standardized way for AI models to communicate with external tools and data sources. Its design revolves around just three primitives:
- Tools: Executable operations exposed by the server. Each tool has a name, description, JSON Schema for input parameters, and permission declarations. The AI model figures out which tool to call and when, based solely on these descriptions.
- Resources: Readable data sources — the equivalent of REST GET endpoints. File contents, database query results, API responses, and real-time sensor data are all unified under this abstraction.
- Prompts: Predefined interaction templates that allow MCP servers to suggest domain-specific best practices to clients.
These three primitives form an interface that is minimal enough to be universal and expressive enough to be useful. MCP doesn't reinvent RPC — it adds an AI-agent-specific semantic layer on top.
The Transport Layer's Three Evolutions
The evolution of MCP's transport layer tells its own story — one of a protocol graduating from a single-machine toy to enterprise infrastructure:
| Phase | Transport | Use Case | Key Limitation |
|---|---|---|---|
| Q4 2024 – Q2 2025 | stdio | Local process communication | Local-only deployment |
| Mid 2025 | SSE (Server-Sent Events) | Remote deployment | Complex connection management, no bi-directional streaming |
| Late 2025 – present | Streamable HTTP | Production | — |
Streamable HTTP is the current recommended standard. It layers streaming capabilities on top of HTTP: long-running tool invocations, server-pushed progress updates, and full compatibility with existing HTTP infrastructure (load balancers, CDNs, firewalls).
The 2025 spec also introduced an OAuth 2.1-based authorization framework, supporting Authorization Code, Client Credentials, and Device Authorization flows — with PKCE mandatory across all flows. This marked the protocol's graduation to enterprise-grade security.
The Three Growth Phases Behind 97 Million
MCP's growth curve can be cleanly divided into three phases, each driven by a fundamentally different force:
Phase 1: Developer Exploration (Q4 2024 – Q2 2025, 0 → 5 million)
Early adopters driven purely by curiosity. Developers integrated MCP into personal toolchains. At this stage, MCP was more of a "cool hack" than a strategic bet.
Phase 2: Platform Integration (Q3 2025 – Q4 2025, 5 → 30 million)
VS Code, Cursor, Windsurf, and other major IDEs and AI coding platforms began shipping native MCP support. Claude Desktop, GitHub Copilot, and Microsoft Copilot all joined the MCP ecosystem. Installs stopped being driven by manual developer configuration — they grew through platform distribution.
Phase 3: Enterprise Adoption (Q1 2026, 30 → 97 million)
The combination of OAuth 2.1 and Streamable HTTP removed enterprise security concerns. OpenAI announced full MCP support. Google Cloud launched MCP Gateway. Major cloud providers began offering managed MCP services. Enterprises started building internal MCP servers at scale — estimated at over 50,000.
The MCP Ecosystem Today
As of late March 2026, the ecosystem numbers are striking:
- Public MCP servers: 15,000+
- MCP packages on NPM: 4,200+
- MCP packages on PyPI: 2,800+
- Enterprise self-hosted MCP servers: estimated 50,000+
- MCP-compatible AI clients: 45+
From Slack to Notion, GitHub to Snowflake, Docker to Postgres — MCP connectors now cover virtually every mainstream enterprise tool stack. Anthropic's February 2026 update announced 50+ MCP connectors, allowing users to read Slack channels, edit Notion docs, and review GitHub PRs without leaving their Claude conversation.
Why MCP Won
Several decisions stand out in retrospect:
Open-source and open from day one. Anthropic didn't build MCP into a walled garden. The protocol spec is public, the SDKs are open-source, and anyone can write an MCP server or client. OpenAI's adoption was the turning point — when a competitor uses your protocol as its core integration interface, it stops being one company's asset and becomes industry consensus.
Gradual complexity, deliberate pacing. The stdio → SSE → Streamable HTTP path gave the ecosystem time to mature at each stage. Shipping Streamable HTTP + OAuth 2.1 on day one would have scared off early developers; staying on stdio forever would have scared off enterprises.
Don't reinvent the wheel. MCP's core design doesn't attempt to replace REST, gRPC, or message queues. It adds an AI-agent-specific semantic layer on top of protocols developers already know: JSON Schema for tool descriptions, OAuth 2.1 for auth, HTTP for transport.
An Operating System for AI Agents?
97 million installs is just a milestone. If MCP continues to penetrate at this rate, it could become the "kernel" of the AI agent era — not the kind that schedules hardware resources, but the protocol layer that lets AI agents interact with arbitrary external systems securely and uniformly.
Just as HTTP standardized information exchange, MCP is standardizing AI agent tool invocation. This may be one of the most consequential infrastructure stories of the late 2020s.