MCP Demystified
MCP is all over my socials today, to the extent that every 4th post is about it. What’s MCP and why should you care? Here I’ll rattle off a bunch of analogies, you can choose what works for you and disregard the rest.
Analogy: API Docs For LLMs
Where it works: Say you have an API that requests a contract draft from Liz every time the API is called. The MCP server tells the LLM how to call your API. It has a name, description, when it should be used, as well as parameters and also general prompt engineering concerns to elicit a reliable tool call.
Where it breaks: MCP also covers the details of how to call your API
Analogy: It’s What the GPT Store Should Have Been
Where it works: Custom GPTs were often used for invoking APIs and tools, but you were limited to one single tool. You would’ve had to open a “Request Contract” GPT in order to invoke your API. With MCP you’d be able to have any chat open and simply connect the “Request Contract” MCP server. In both cases, the LLM is still responsible for invoking your API. It’s dramatically better, because now the LLM can use all your APIs.
Where it breaks: It’s pretty good. It’s a different paradigm and a lot more technical, so many people probably don’t vibe with it.
Analogy: LSP (Language Server Protocol) for LLMs
Where it works: LSP & MCP both solve the many-to-many problem. For LSP it’s IDEs vs programming languages. For MCP it’s LLM clients (e.g. ChatGPT, Cursor or an agent) vs tools/APIs/applications/integrations.
Where it breaks: It’s pretty good. The actual integrations feel a bit more fluid in MCP because so much of it is natural language, but that’s the essence.
Analogy: Power Tools for AI
Where it works: Power tools have a lot of standard interfaces, like you can put any drill bit into any drill. Also, many power tools have very similar user interfaces, e.g. a hand drill and a circular saw both have a trigger.
Where it breaks: This one feels like a bit of a stretch, but it does convey a sense of being able to combine many tools to complete a job, which is good.
MCP Server Ideas
There are a lot of existing MCP servers, including, Gitub, Google Maps, Slack, Spotify (play a song), PostgreSQL (query the database), and Salesforce. Some others that could be:
- Browser use (load a page & click around)
- Microsoft 365 (I’d love to get an org chart in an LLM)
- Wikis & documentation
- YouTube
- Email (mainly searching & reading, but also maybe sending, 🤔 maybe)
FAQ: How do I integrate MCP into my enterprise?
You would choose a LLM chat tool that supports MCP and then configure and connect MCP servers. I’d imagine you’d want to connect your wiki, Salesforce, maybe a few CRM systems. At the moment, heavy enterprise integration would require your IT department slinging some code to build MCP servers.
It’s an Anthropic project, so Anthropic tools all have great support, whereas OpenAI and Microsoft are going to shun it for as long as possible. But servers are easy to create, expect community servers to pop up.
FAQ: Why?
Universal integrations into AI. All you have to do to get your company into the buzz is wrap your API in an MCP server, and suddenly your app can be used by all MCP clients (Claude, Cursor, agents, etc.)
FAQ: What if BIGCO X develops a cometitor? Who will win?
The one that has more users. It’s a protocol. Which is better has little to do with it, it’s all about which has the biggest network effects. I’d bet on MCP because it was released months ago and there’s a ton of buzz around it still.
FAQ: IDK it still seems hard
Okay, maybe a diagram helps
Servers on left; clients on right. Redraw the arrows however you want.