GitHub has published a short piece that takes five confident sentences about AI and asks under what conditions each is true. In Should you read the code, is RAG dead, and did Skills kill MCP?, the framing is the useful part: “Hot takes turn complicated topics into one confident sentence. That makes them great for engagement, but not necessarily for understanding.” On reading AI-generated code, the answer is a rule rather than a slogan: “review until you can explain and own the outcome,” because “AI moves the effort around. It does not make the work disappear.” On RAG: “RAG is not dead. It is just not the newest thing people want to post about.”
There are no figures in the post, vendor or otherwise, so there is nothing to discount. The take with the most at stake for API providers is the third. “Skills killed MCP.” “No. They solve different problems.” The distinction is drawn in two sentences that deserve to be the standard version: “MCP can provide access. Skills can explain how to use that access well.” A skill is “packaged expertise,” often Markdown, which means “people can read them too,” while MCP is the structured interface an agent calls. And then the instruction: “Use standards for shared interfaces. Use skills for context, process, and best practices.” It is the same conclusion TestMu AI reached by measurement two weeks ago in this series, arrived at here from the practitioner side.
This is a claim the catalog can test against the company making it. The GitHub provider page lists 56 API pages, and the ones this conversation runs through are the GitHub Copilot API and the GitHub Models API. On the Agent Readiness score both dimensions the post discusses are lit: the MCP server and agent skills. GitHub is not arguing for a theory. It ships both layers, and the record shows them side by side. The agentic access profile maps 2,431 operations, 1,203 of them acting and 31 flagged human-in-the-loop.
The Kin Score is 80.3, exemplar band, carried by operational transparency at 92.1 and access clarity at 84.2, with contract quality at 75.6 and contract governance at 31.8. The Agent Readiness score is 59.9, agent-ready, with delegated identity and protected resource metadata lit alongside the two layers the post defends. What holds it below agent-native is idempotency, which is unlit, and dynamic client registration and the well-known catalog are unlit as well. That is worth a sentence in a post about reviewing until you can own the outcome. An agent with access through MCP and method through a skill still has 1,203 ways to change something, and the contract does not say which of them are safe to repeat. The hot take was wrong and the correction is right. The harder question is the one no hot take asks.