beingAInative
Subscribe

Answers · Connecting AI to your tools

RAG vs MCP: which one do you actually need?

The short answer

RAG vs MCP is comparing a technique to a standard, which is why the debate never resolves and probably never will.

RAG, retrieval augmented generation, is a technique: before the model answers you, your app searches your documents for relevant passages and stuffs them into the prompt, so the answer draws on your material instead of just the model's training. MCP, the Model Context Protocol, is an open standard Anthropic released in November 2024 that defines how a model finds and uses external tools and data.

So RAG is one thing you might build. MCP is how a model reaches whatever you built. A search tool behind an MCP server? That's both, working together.

August 16, 2026 · Subscribe for more groovy writing

People argue about this one like it’s a cage match. It’s not. Let me show you why the question itself is a little bit broken, and then give you the decision rule I actually use.

When is RAG the answer?

When you have a pile of documents and you want answers grounded in them. Your knowledge base. Past client notes. A course library. Three years of newsletters. RAG is the shape of that solution no matter which protocol carries it.

Now, the quality of a RAG setup lives in genuinely unglamorous places (how the documents get chunked, how search ranks the passages, whether the model actually cites what it retrieved). Nobody puts that part in the demo.

But here’s the shortcut most people miss: if you use Claude Projects or ChatGPT with uploaded files, congratulations, you’re already using RAG someone else built. That covers more use cases than you’d think, and it costs nothing to try before you build anything.

When is MCP the answer?

When you want the model to do things, or to reach data that lives behind an interface instead of inside documents.

Read your project board. Query your database. Pull this month’s orders. Those are tools, and MCP is the standard plug that lets Claude or another assistant find and call them without custom wiring per tool. (I broke down how that plug relates to the APIs underneath it in the MCP vs API answer.)

A machine console whose screen asks what the model is missing: a bookshelf slides a book along an arrow labeled RAG into the console's wide flat slot, and an open toolbox sends a wrench along an arrow labeled MCP into its round tool port.

The decision rule

Ask what the model is missing. That’s it, that’s the rule.

Missing knowledge from your documents? That points to RAG, and the built-in version in your chat tool is the right first stop. Missing hands, the ability to act or see live data? That points to MCP.

And a growing number of real setups are honestly both at once: an MCP server whose tool happens to be a retriever. The plumbing argument matters way less than knowing which gap you’re closing.