Every few months a new acronym sweeps through the AI world and everyone has to pretend they understood it from day one. MCP — Model Context Protocol — is the latest. The good news: this one is genuinely simple, and it's the most useful thing to happen to consumer AI since the chat box.
If you've ever wished your AI assistant could actually do something in the apps you use every day — see what's in your fridge, check your calendar, update your shopping list — that's what MCP unlocks. Here's the plain-English version, with no hand-waving and no "in today's rapidly evolving landscape".
The browser-extension analogy
The clearest way to understand MCP is to think about browser extensions.
Before extensions, your browser was a closed box. It rendered web pages and that was it. Then Chrome (and Firefox before it) added an extension API, and suddenly anyone could write a small piece of software that plugged into the browser to add new capabilities — password managers, ad blockers, screenshot tools, the lot. The browser became a platform.
MCP is the same idea, but for AI assistants. Before MCP, Claude and ChatGPT were closed boxes too. They could chat with you, but they couldn't reach out to anything in your life. Now any developer can write a small piece of software — an "MCP server" — that plugs into the AI and gives it new powers. Read your pantry. Update your meal plan. Send an invoice. Run a database query. The AI becomes a platform.
Browser extensions made the browser useful for more than reading. MCP apps make AI assistants useful for more than chatting.
What does "Model Context Protocol" actually mean?
Three words, three jobs:
- Model — the AI model (Claude, ChatGPT, Gemini, etc.). The thing doing the thinking and the talking.
- Context— everything the model needs to do its job that isn't already in the conversation. Your data. Your tools. Your history.
- Protocol — a standard, open way of getting that context into the model. Defined once, works everywhere.
Anthropic published the spec in late 2024 and made it open source. Within a few months, every major AI client started supporting it. As of 2026, you can use the same MCP server inside Claude, ChatGPT, Cursor, Cline, Windsurf, and a long tail of other AI clients. Build it once, it works everywhere — exactly like a web standard.
So what's an "MCP app"?
An MCP app is a piece of software that exposes tools and resources to AI assistants over the MCP protocol.
- Tools are things the AI can do. "Add eggs to my shopping list." "Mark this meal as cooked." "Search my food history for that lasagne we had." The AI calls them like functions, the MCP app does the work, the AI sees the result.
- Resources are things the AI can read. Your pantry contents. Your family's dietary restrictions. The text of a recipe. The MCP app fetches them on demand and the AI uses them to answer your question.
Some MCP apps also include UI widgets — small embedded interfaces that render inside the AI client itself. Think of them as the visual equivalent of a browser extension popup. Surprise Chef uses them to render proper recipe cards and pantry audit walkthroughs without making you leave your AI conversation.
Why this matters more than people think
Until MCP, the way most apps integrated AI was: bake a model into the app, charge users a recurring subscription for the privilege, hope they don't notice they're paying twice (once for ChatGPT Plus, once for the app). That's the dominant pattern in 2025 and most of 2026.
MCP inverts it. Now apps can stop pretending to be AI companies and just be apps again — they provide the data, the tools, the domain expertise — and the user's existing AI subscription provides the brain. Better for users (no double-billing), better for app builders (no per-token cost to manage), better for AI providers (their assistants get more useful with every new MCP app shipped).
It's also better for continuity. A bundled-AI app forgets everything between sessions because it has nowhere to put state. An MCP app keeps the state on its own server, so the AI gets fresh, real data every time. "What did we eat last week?" actually works, because the food history isn't living inside a chat thread that got summarised away.
What makes a good MCP app?
Having shipped one (and used a lot of others), three things consistently separate the great MCP apps from the merely working ones:
1. Tools that name what they do, not how they work
The AI is going to read your tool descriptions and decide whether to call them. "get_pantry— Returns the current pantry contents for the household" is a tool. "fetch_v3_pantry_with_joins" is a leaky abstraction. The first one will get used. The second one will get ignored.
2. State that lives outside the chat
Anything that should survive past the current conversation needs to be on the MCP server, not in the chat history. Surprise Chef puts your meal plans in a database. A note-taking MCP app puts your notes in its own store. A good rule of thumb: if a user closes the chat and opens a new one tomorrow, what should still be there? That's what you persist.
3. Honest scope
The temptation when building an MCP app is to expose 50 tools because "the AI can figure out which one to use." It can't, reliably. Better to ship 10 tools that do one job each, and let the AI compose them. Cold-model evaluation — testing your app with a fresh AI session that has no priors — is the only way to find out which tools survive.
An example: Surprise Chef
Since we built it, here's the elevator pitch for our own MCP app: Surprise Chef gives your AI persistent memory of your kitchen. Pantry, family dietary profiles, meal plans, food history, recipes — all of it lives on a server you own, all of it is reachable from any MCP-compatible AI client.
You install it once, connect it to your Claude or ChatGPT account, and from then on you can ask your AI anything kitchen-related and it will actually know the answer. A few real prompts:
- "Plan my week of meals around what's in my pantry."
- "What can I cook tonight with what I have?"
- "Save this cookbook page as a smart recipe." (paste a photo)
- "Find a recipe like that lasagne we had three weeks ago."
None of those work in a vanilla AI chat. All of them work the moment Surprise Chef is connected, because the AI suddenly has tools and resources for your real kitchen. That's what an MCP app feels like in practice.
How to try one
If you're ready to actually use MCP rather than just read about it, here's the shortest path:
- Sign up for a free Surprise Chef account — name your kitchen, takes about a minute.
- From the dashboard, follow the Claude or ChatGPT setup guide. About 90 seconds.
- Open a new chat in your AI client and ask "what can I cook tonight?" — the answer will come from your real pantry, not generic web data.
That's the whole point of MCP, in three steps. Once you've used one good MCP app, every chatbot without one starts to feel weirdly amnesiac.