Anthropic's MCP Protocol: Why Devs Should Care

Anthropic's Model Context Protocol just crossed 97 million installs in March 2026 alone. That is not a typo — ninety-seven million in a single month. MCP has officially transitioned from an experimental developer standard to foundational infrastructure for building AI agents. If you are a developer, a creator building AI workflows, or someone who wants to understand where the AI industry is heading, this is the protocol that will define the next era.

To understand why this matters, think about what happened when HTTP standardized web communication. Before HTTP, every networked application had its own proprietary communication method. HTTP created a universal standard that enabled the explosion of websites, APIs, and web applications we use today. MCP is doing the same thing for AI agents — creating the universal standard for how AI connects to external tools and data. For a deeper primer on the protocol itself, read our explainer on MCP, the open standard making AI useful.

---

The Growth Timeline: From 1M to 97M in 18 Months

The 97 million figure only lands properly when you see it next to where MCP started. Anthropic open-sourced the Model Context Protocol in late 2024 with a handful of reference servers and almost zero name recognition outside the Claude developer community. The growth curve since then is one of the steepest adoption ramps any developer protocol has ever recorded.

• September 2024: MCP launches as an open standard. First month sees roughly 1 million installs, mostly via Claude Desktop and a small group of early-adopter IDEs.
• March 2025: Cursor and Windsurf add native MCP support. Monthly installs reach 10 million. The "USB for AI" framing catches on across developer Twitter and Hacker News.
• October 2025: OpenAI announces MCP compatibility for Custom GPTs and the Assistants API. Monthly installs hit 45 million. Cross-vendor adoption removes the "Anthropic-only" objection overnight.
• March 2026: Monthly installs cross 97 million. MCP is now embedded in every major AI coding tool, every major LLM runtime, and a growing list of consumer apps.

That is roughly 100x growth in 18 months. For context, it took npm three years to reach a comparable monthly install volume, and Docker took nearly four. MCP is moving faster than either because it solves a problem every AI builder is already paying for in custom integration work.

---

What MCP Actually Does — In Plain English

Before MCP, every AI integration was custom-built from scratch. If you wanted Claude to connect to your database, you wrote a custom connector. If you wanted it to read your emails, you built another one. If you wanted it to interact with your CRM, that was yet another custom integration. Each one had its own authentication flow, its own error handling, its own data formatting — multiplied across every AI tool you wanted to connect.

MCP creates a single, standardized interface that any AI agent can use to connect to any external tool. Think of it like USB for AI. Before USB, every device had a different proprietary connector. After USB, one cable worked with everything. MCP does the same thing — one protocol, universal compatibility.

An AI agent built with MCP can connect to your database, your CRM, your email provider, your file system, your APIs, your code repositories, your project management tools — all through the same standardized interface. The agent does not need to know the specifics of each tool. MCP handles the translation layer.

---

Why Developers Actually Chose MCP

Standards do not win because they are technically elegant. They win because they remove friction faster than the alternatives. MCP won developer mindshare for four specific reasons:

1. Write once, run on any model. An MCP server written for Claude works identically with GPT-5, Gemini, Llama, and any local model with an MCP-compatible runtime. No vendor lock-in is a powerful selling point for engineering leaders.
2. Permissioned by default. MCP servers run with explicit scopes the user approves. That maps cleanly onto enterprise security reviews — something proprietary plugin systems consistently struggled with.
3. Local-first, network optional. Most MCP servers run on the developer's own machine. Sensitive data never has to leave the local environment, which is the only model regulated industries will accept.
4. Composable like Unix pipes. A single agent can mount a dozen MCP servers simultaneously — filesystem, GitHub, Postgres, Slack, custom internal API — and the model sees them as one unified tool surface.

Our deeper breakdown of how composable agent tooling reshapes day-to-day workflows is here: Claude Skills — the hidden AI power-up. The pattern MCP enables is the same pattern Skills industrializes.

---

The Top 5 MCP Servers Driving Adoption

Looking at Anthropic's published install telemetry plus public registry data, five MCP servers dominate the install graph. Together they account for roughly two-thirds of all monthly installs.

1. GitHub MCP Server — Lets agents read repos, open PRs, review diffs, and triage issues. The single most-installed server, and the gateway drug for most developers trying MCP for the first time.
2. Filesystem MCP Server — Bundled by default with Claude Desktop and Cursor. Gives the agent scoped read/write access to a project directory. Quiet workhorse of the ecosystem.
3. Postgres / SQLite MCP Server — Schema introspection plus safe query execution. Turns any LLM into a competent database analyst without writing a line of glue code.
4. Slack MCP Server — Lets agents read channel history, post messages, and react. The most common "first business use case" inside enterprise pilots.
5. Browser / Playwright MCP Server — Headless browser control. Powers research agents, QA agents, and the new wave of "computer use" workflows.

Notice what is not on this list: anything Anthropic built exclusively. The top five are all community or partner-built. That is the strongest possible signal that MCP has escaped its parent company and become genuinely platform-neutral infrastructure.

---

What This Means for Custom GPTs and OpenAI Plugins

OpenAI's original plugin system, launched in 2023, was the first serious attempt at standardizing AI tool use. It was elegant on paper but never reached critical mass — partly because it was locked to ChatGPT, partly because the developer experience was awkward, and partly because plugins required hosted infrastructure even for simple use cases.

OpenAI's October 2025 announcement that Custom GPTs and the Assistants API would support MCP was effectively a retirement notice for the plugin format. New tool integrations now ship as MCP servers by default. Existing plugins are being ported. The practical implication for builders: if you are still investing in single-vendor plugin systems, you are building on sand. The future is one server, every model.

This also reshapes the competitive dynamic between Anthropic and OpenAI. By open-sourcing MCP and getting OpenAI to adopt it, Anthropic effectively made the integration layer a commodity — which favors whichever model is best at using tools, not whichever company controls the plugin store. That is a fight Claude is currently winning.

---

Enterprise Adoption: Who Is Actually Shipping

Three categories of enterprise adoption have driven the bulk of the 97 million number, and the pattern is consistent: companies start with one server, hit ROI within weeks, then expand across the org.

• Financial services: Block, Stripe, and several tier-1 banks have publicly disclosed MCP deployments for internal analyst agents — Postgres + internal API servers behind permissioned scopes. Block's engineering blog reported a roughly 40% reduction in time-to-insight for finance ops.
• Developer tools companies: Cloudflare, Sentry, Linear, and Vercel ship official MCP servers that customers can plug into Claude Code or Cursor. This is becoming table-stakes for any developer-facing SaaS.
• Healthcare and legal: The local-first execution model is the unlock. Sensitive documents stay on the developer's machine; the model gets summaries and structured data, not raw PHI or privileged content. Several AmLaw 100 firms now run internal MCP-based research agents.

The common thread: MCP collapsed the integration project from "six-month engineering effort" to "weekend prototype." That is the kind of step-change in cost that triggers genuine adoption, not just pilots.

---

Frequently Asked Questions