MCP is suddenly everywhere
A week-of note on Model Context Protocol finally hitting critical mass.
Anthropic announced MCP back in November and for about three months it looked like it might quietly die. Now it’s suddenly in every tool I use.
Quick version of what it is, in case you’re reading this in 2027 and have forgotten the history: Model Context Protocol is an open spec for how LLM clients talk to external tools and data sources. Instead of every app inventing its own way to expose a database or a filesystem or a Slack workspace to an LLM, you run an MCP server that exposes those things once, and any MCP-aware client can use them. Think LSP but for AI tools instead of editors.
Why it took a few months to catch: the original launch shipped a spec and a handful of reference servers, but the clients were almost entirely Claude Desktop. Without clients, servers are useless. Without servers, clients are useless. Classic chicken-and-egg.
What changed this month: Cursor added MCP support. Zed added it. There are a couple hundred community servers on GitHub now, including ones for Postgres, GitHub, Linear, Sentry, Obsidian, and basically every SaaS tool a developer touches. Once you have a critical mass of servers, every new client benefits immediately. Once you have a critical mass of clients, every new server benefits immediately. The flywheel started spinning around last week and I don’t think it’s stopping.
The thing I want to note, for future-me: this is the first time we’ve had a real interop layer for AI tooling. Every previous attempt was one vendor’s SDK. MCP is going to look in hindsight like the moment AI tools stopped being silos.
The other thing I want to note: I was wrong about it in December. I thought it was going to be another well-intentioned spec nobody adopted. I should be more patient with protocol launches — they always look dead for the first 90 days.