RAG Is Not Dead, MCP Is Not Killed: Testing Hype Claims

Last week, GitHub Blog examined three fashionable claims from developer Twitter at once: there is no need to read code, RAG has outlived its usefulness, and Skills finished off MCP.

Each sounds convincing in a single sentence and falls apart under the first test.

The company pays for AI tools with real budget.

For her, the difference between a striking claim and a working solution is the difference between an investment and a write-off.

The GitHub Blog author realized that hype claims are a convenient format for engagement but useless for decision-making.

The reader agrees, argues for five minutes, and scrolls on.

A reader benefits from a claim only when breaking it down into conditions: when it is true, what is missing from the wording, and what changes when it is applied to a real task. This analysis is especially clear with RAG and MCP because both claims concern real technology but describe it inaccurately.

RAG is a mechanism: the model searches a vector database for relevant fragments, adds them to the prompt, and only then responds.

It does not replace an agent with access to live APIs, and the claim is partly right here: where data can be obtained through a direct tool call, an extra retrieval layer is genuinely redundant.

But when the source is a private database containing thousands of documents—regulations, a 1C catalog, or product taxonomy in Pimcore or Akeneo—RAG remains the only way to give the model precise context without retraining.

The claim “RAG is dead” confuses a specific case with a general rule.