Double click: What does MCP mean for agentic AI?

Anthropic introduced the Model Context Protocol (MCP) back in November, but it was only after OpenAI announced its support for MCP earlier this year that it hit peak virality. In a nutshell, MCP creates a common protocol for how AI assistants like Claude, Copilot, or Cursor communicate with external tools and data sources, so developers don’t have to create custom integrations for each one. You’ve probably heard it being compared to a USB-C port for AI applications, or as Microsoft Chief Technology Officer Kevin Scott put it at this year’s Microsoft Build conference, “akin to HTTP for the internet, where it allows you to do sophisticated things because the protocol itself doesn’t have much of an opinion about the payload that it carries.”

By giving large language models (LLMs) a way to connect to tools and data sources in real time, MCP holds huge promise for agentic AI and its ability to act on our behalf. Since it hit the scene, the list of available servers grows daily. In fact, we’ve built our own Dev Mode MCP server to bring Figma into developer workflows and help LLMs generate design-informed code. So what does this new standard say about where we’re headed?

MCP is an accelerant

You might be wondering, “How is using an MCP server to generate code with an LLM better than using one that connects to an API?” You’re not alone: Pat Walls , Founder at Starter Story, posed this question on X. “With MCP, LLMs are aware of various ‘tools’ and will interact with them mid-conversation to expand their capabilities,” answered user @LostandFounding. “In your example, the LLM would have to generate the code and then run the code every time you want to interact with a tool—would be really impractical and inefficient.” With MCP, you get speed and scalability, which means that now more than ever, you can just do things.

Shubham Saboo , Head of Developer Relations at Tenstorrent, built the “automated AI travel agency” many of us have envisioned with four agents working across Google Maps, Airbnb, Google Calendar, and Weather. Siddharth Ahuja , Co-founder of Prophecy, created a Blender MCP server to create a 3D scene of a “low-poly dragon guarding treasure” with just a few sentences. And recently, Y Combinator hosted “the world’s biggest MCP hackathon” with the winners Rajmeet Singh, Rishabh Chanana , and Chris Hailey taking first place with Observee, an observability platform for MCP calls. 

Quality in, quality out

It’s not just about doing things, though—it’s about doing them well. John Kutay , Director of Engineering at Rippling, wrote on X that the protocol can help enrich AI: “MCP, when implemented well, delivers the determinism AI needs for business apps…it can use AI in clever ways rather than being a slop machine.”

Jake A. , Developer Advocate at Figma, underscores the importance of precise contexts: “On their own, MCP servers extend access to context, but determining the best context to provide through those new channels is still very much where expertise lies.” As he explains with the beta release of Figma’s Dev Mode MCP server, the better alignment there is between design and code, the better the results of the server. “The MCP server is a multiplier—not a replacement—for alignment,” he says.

As we’ve heard, efficiency doesn’t mean we leave craft behind. Quite the opposite: It means we become more discerning and precise when it comes to our process and output.

***

To learn more about how MCP is laying the foundation for an open agentic web and why widespread adoption matters, read the full article.

Like what you see? Sign up to get special editions of The Long & Short of It delivered straight to your inbox.