In March 2026, Model Context Protocol crossed 97 million monthly SDK downloads. It now has over 10,000 public servers across registries, with official SDKs in TypeScript, Python, C#, Java, and Swift. Anthropic, OpenAI, Google DeepMind, Microsoft, Salesforce, Block, Cloudflare, and Replit all adopted it — within months of each other.
MCP is now a top-five technology priority for most software engineering organisations. Security is the leading blocker. Adoption is accelerating anyway.
And yet most developers who aren't deep in the AI tooling world still think of it as a niche Anthropic thing — something to do with Claude Desktop. That understanding is several months out of date, and the gap is going to matter.
What MCP Actually Is
MCP stands for Model Context Protocol. It's an open standard that defines how AI agents communicate with external tools and data sources. Before MCP, every AI integration was bespoke — you'd write a custom connector for each tool, each agent, each workflow. The connection between your AI and your database, your GitHub, your Slack, your CRM was a one-off glue script.
MCP replaces that with a standard interface. You build an MCP server for your tool once. Any MCP-compatible AI client can then discover and use it — without custom integration code. It's the difference between every device having its own proprietary charger and USB-C existing.
The analogy that actually holds
HTTP standardised how web clients and servers communicate. REST standardised how APIs are structured. MCP is doing the same thing for AI-tool communication — defining the protocol layer so agents and tools can talk without custom wiring.
An MCP server exposes three things: tools (callable functions), resources (data the agent can read), and prompts (reusable templates). The agent discovers what's available, calls what it needs, and gets structured responses back. No scraping, no guessing, no brittle UI automation.
Why Cross-Provider Adoption Is the Signal
Standards only become infrastructure when all major players adopt them. That's what makes the MCP story significant — it isn't just Anthropic's protocol anymore. OpenAI integrated it into the ChatGPT desktop app. Google DeepMind adopted it. Microsoft built it into Semantic Kernel and Azure OpenAI. Salesforce shipped it in Agentforce.
The defining characteristic of infrastructure standards is that they win by becoming invisible. Nobody thinks about TCP/IP when they load a webpage. Nobody thinks about HTTP when they call an API. MCP is on that trajectory — the thing that makes agent-tool communication work without anyone having to think about it.
What that means practically: in 18 months, the question won't be 'should we support MCP?' It'll be 'why don't you have an MCP server yet?' The same way not having an API was a liability a decade ago.
What's Already on MCP
The ecosystem moved fast. Official MCP servers now exist for GitHub, Slack, Google Drive, PostgreSQL, Notion, Jira, Salesforce, and hundreds of other tools. Most major SaaS platforms either ship one or have a community-built version.
- GitHub MCP server: agents can read repos, create PRs, check CI status, post comments — without browser automation
- PostgreSQL MCP server: agents can query databases with structured, auditable access
- Slack MCP server: agents can read channels, post messages, search history
- Notion MCP server: agents can read and write pages, search workspaces
- Custom servers: any developer can build and publish one — the SDK is straightforward
The 10,000-server figure is a registry count from early 2026. By the time you read this, it will be higher.
What This Means If You're Building Something
If you're building a product that developers or AI agents will interact with, the calculus is simple: ship an MCP server. Not because it's technically interesting — because it's becoming table stakes for agent-ready software.
If you're building agents or internal AI tooling, MCP is the reason you don't need to write custom connectors for every tool you want your agent to use. The plumbing already exists. You connect to it.
The security concerns are real and worth taking seriously — agents with broad MCP access can do significant damage if misconfigured or compromised. That's not a reason to avoid MCP; it's a reason to implement it carefully, with scoped permissions and audit logging.
“Infrastructure standards only become infrastructure when all major players adopt them. MCP crossed that threshold in early 2026. Everything after this is just catching up.”
Most developers are still in the 'catching up' phase. That's not a criticism — the adoption happened fast. But if you're building anything that touches AI agents, the time to understand MCP is now, not when your competitors already have servers in the registry.