What is MCP in plain English?
MCP is a standard way for an AI assistant to plug into other software. Instead of every AI app writing its own connector to every product, a product publishes one MCP server, and any AI app that speaks MCP can use it.
Think of it as a universal port. Each product exposes its features once, in the same format, and the assistant discovers what it can do. Anthropic introduced it as an open protocol, with the specification at modelcontextprotocol.io. It sits underneath much of our hub on AI product UX, because it's how chat and agent features reach real data.
How do MCP servers, clients and hosts work?
MCP has three roles. The host is the AI app you use, such as a chat app or a coding editor. Inside it, a client keeps a connection to each server, and each server wraps one product or data source.
In the official specification, a server can offer three kinds of features:
- Tools: actions the model can call, such as "create a page" or "get this frame".
- Resources: data the model can read as context, such as documents or records.
- Prompts: ready-made templates a user can choose, such as "summarize this file".
Servers run locally or remotely, and remote servers usually ask you to sign in, so the agent acts with your account's access. The spec says hosts must get user consent before a tool runs, and it's revised in dated versions, so check modelcontextprotocol.io for current details. Our AI infrastructure industry page shows how we explain products like this.
How does MCP compare to APIs and plugins?
MCP doesn't replace an API. Most MCP servers are a thin layer over a product's existing API, describing its actions so a model can discover and call them. A plugin is usually built for one specific app.
Term | Who uses it | What it does |
|---|---|---|
API | Developers writing code | Exposes a product's data and actions to programs |
Plugin or extension | Users of one specific app | Adds features to that one app |
MCP server | Any MCP-compatible AI app | Describes a product's actions and data so a model can use them |
AI agent | End users | Plans and runs steps, often by calling MCP tools |
An API is written for a developer who reads docs; an MCP server is read by a model at runtime, so clear tool names become part of your UX. Our guide to agentic UX covers what happens when an agent chains those tools.
What does MCP look like for designers and founders?
For designers, MCP lets a coding agent read a Figma frame's layers, variables and components instead of guessing from an image. For founders, it lets an assistant pull from your CMS, docs or database.
Hypothetical example: an agent reading a Figma file and a CMS
For example, suppose a founder wants a new pricing section. Their coding editor has two MCP servers connected: Figma's and their CMS's.
They paste a link to the pricing frame. The agent calls a Figma tool for the frame's structure and variables, then a CMS tool for current plan names and prices, and writes the section with existing components. Before any CMS tool that would publish, the editor asks for approval. The founder reviews the diff, fixes mobile spacing and ships it.
Our guides to the Figma MCP server and Figma to code with AI cover where a developer still steps in.
What are common misconceptions about MCP?
The most common mistake is treating MCP as a feature users see. It's plumbing. Users see approvals, connected accounts and tool activity, and those need design.
- "MCP makes AI accurate." It gives the model better context, but the model can still misread data or call the wrong tool.
- "An MCP server is secure by default." The server only has the access you grant. Scopes, consent and revoking access are your job to design.
- "Every client supports everything." Support for MCP features varies by client and changes often, so check each client's docs.
