Nick Cooper · 02:14 "A protocol is just a language."
In May 2026, with thousands of MCPs already proliferating, the industry's binary debate of "MCP vs CLI — which should you use?" finally got an explicit answer from OpenAI: not either/or — use both.
The speaker is Nick Cooper, a technical staff member at OpenAI who works on MCP (Model Context Protocol), and one of the core maintainers. He's also a board member of the Agentic AI Foundation (AAIF), launched under the Linux Foundation in January 2026, and was central to OpenAI's contribution of AGENTS.md to AAIF. Early in the talk he mentions in passing that "I used to teach computer science in high school" — a background that comes through in how he reframes protocols as "just a language" rather than adversarial jargon.
The entry point is the history of layered structure. Electricity / light → TCP → HTTP → REST → OpenAPI → MCP — a chain of language evolution in which each layer added new verbs and structure. MCP is the topmost layer, but ultimately just the next link in the same lineage.
The central claim is that "MCP is an API for AI" (10:18). The message: don't just mechanically convert existing APIs into MCP — design intent on the assumption that AI models will use it. The example is an email server. The API looks simple — "send and receive mail" — but once you drop into the implementation there are hundreds of verbs: labels, folders, drafts, contacts, classifications. Show all of them to an AI model and it burns through context before the problem is even stated. What the user (the AI) actually wants is just two things: "write a draft" and "send it." The high-level intent needs to be narrowed.
The answer to the title's question — "MCP vs CLI" — lies in a shift of design philosophy: not transforming an existing API at the interpretation layer, but placing intent at the center. CLI has decades of accumulated tooling on the shelf; MCP is the layer that carries the AI's intent. The two don't compete — strong systems combine them.
Key Observations
The reframe: a protocol is just a language (02:09)
"Protocol" sounds specialized and forbidding, but ultimately it's just a shared language systems use to talk. The history — electricity → TCP → HTTP → REST → OpenAPI → MCP — can be relined up as "language evolving by gaining more verbs and more structure." The framing feels fresh. The ex-high-school-teacher's habit of de-specializing the jargon sets the tone of the talk.
A simple API expands into hundreds of tools (10:43)
When turning an email server into MCP, the existing API has hundreds of verbs — add label, save draft, fetch contacts, classify, create folder. Show all of them to an AI model and "it uses up its context before the problem is stated." What the user (the AI) wants is just two things: "write a draft" and "send the email." An API design narrowed to high-level intent is OpenAI's MCP design guideline. Tokens are more expensive on output and take more time, so input/output asymmetry has to be built into the design from the start.
Code mode as a third option (06:39)
Two patterns are evolving — the shell line (file-system-like abstraction; resources are files; tools are executables) and the code line (functions in TypeScript / Python). Bash is also a proper programming language, and can even handle asynchronous work. Rather than instructing the model "run A, then B, then C" step by step, expressing it as a single program is more token-efficient. "Not using shell or bash, but writing it" — a third paradigm.
Video outline
- (00:00) Self-introduction, framed as a continuation of last year's MCP talk
- (02:09) Protocols are just a language — the layered structure from electricity to MCP
- (03:51) Each layer creates standardization and specialization — injection points for auth, security, telemetry
- (04:24) The asymmetry of token efficiency — output costs more than input and takes more time
- (05:32) The MCP explosion — the "context exhaustion" problem with thousands of MCPs proliferating
- (06:11) The two patterns under evolution — shell-like (file-system-style) and code-like
- (07:48) Code mode — Bash is also a proper programming language
- (11:08) Answer to the MCP vs CLI debate — not either/or, use both
- (12:35) Next horizon — MCP as an inter-agent protocol
- (13:24) Identity and discovery — mechanisms to discover the flood of MCPs are needed
Sources
MCP vs CLI OpenAI Engineer Says You Need Both — Agentic AI Foundation