A Bold Claim: MCP Is Already Dying

Developer Eric Holmes has published a provocative piece arguing that the Model Context Protocol (MCP), despite the industry's initial enthusiasm, is already on its way out. His core argument: LLMs are already excellent at using CLI tools, and MCP adds unnecessary complexity on top of something that already works well.

LLMs Already Understand CLIs

LLMs have been trained on millions of man pages, Stack Overflow answers, and GitHub repositories full of shell scripts. When you tell Claude to use gh pr view 123, it just works. MCP promised a cleaner interface, but Holmes found himself writing the same documentation anyway: what each tool does, what parameters it accepts, and when to use it.

Three Core Advantages of CLI

Debuggability: When Claude does something unexpected with Jira, you can run the same jira issue view command and see exactly what it saw. With MCP, you are spelunking through JSON transport logs.
Composability: CLIs compose beautifully. terraform show -json plan.out | jq '[...]' is practical CLI usage that MCP cannot easily replicate without custom filtering servers.
Auth compatibility: aws sso login, gh auth login, kubectl config — battle-tested auth flows that work the same whether a human or Claude is driving.

Counterarguments and Caveats

MCP does have legitimate use cases: services without CLI interfaces, GUI-only tools, and scenarios requiring standardized interfaces across multiple LLMs. But Holmes' argument lands squarely in the majority case: if a CLI already exists, building an MCP server on top is over-engineering.

Community Reception

The post earned 234 points on Hacker News and sparked a rich debate. The debate ultimately reflects a broader tension in the AI tooling ecosystem: protocol standardization vs. embracing the existing Unix philosophy.

#ai-tooling

When Does MCP Make Sense vs CLI?

When Does MCP Make Sense vs CLI?