CLI vs MCP: How AI Agents Choose the Right Tool for the Job

Summary

The video compares two ways AI agents interact with software and external services:

• CLI (command-line interface): The agent runs ordinary terminal commands, such as cat, grep, git, and curl, much like a developer would. Models often already know common commands and flags from their training, so they may not need extra instructions or tool definitions.
• MCP (Model Context Protocol): MCP servers provide named, structured tools, each with a description and a JSON schema specifying its inputs. A server can also handle work—such as authentication or rendering a webpage—that would otherwise fall to the agent.

Demonstrations and findings

Reading and searching files

• For a task to read notes.md and find the word “agent” across Markdown files, the agent used cat and grep with CLI.
• With MCP, it used the file-system server’s read file and search files tools.
• Both methods succeeded. CLI was more compact; the MCP server exposed 13 tools, including 11 that were unnecessary for the task. Their schemas still took up context.

Git operations

• The presenter says an agent can use familiar Git commands, such as checking recent commits or working-tree status.
• The GitHub MCP server in the example exposed 80 tools. The presenter estimates their definitions consume about 55,000 tokens, even when a task needs only one or two tools.
• This illustrates the CLI argument: for common local developer tasks, MCP schemas can use context and incur token costs for knowledge the model may already have.

Fetching a JavaScript-rendered webpage

• Using a Fetcher MCP server, the agent made one fetch URL call. The server used a headless browser to render the page and return readable text; the presenter reports that this took a few seconds and used about 250 tokens.
• With CLI, curl returned mostly JavaScript because the site was a Next.js application. The agent tried text processing and then wrote a Python script to inspect the page’s data format. The presenter says it took several minutes and more than 2,000 tokens to retrieve enough content.
• This shows MCP can be useful when a server provides an abstraction beyond what a raw command returns.

Takeaways

• CLI tends to suit familiar, direct tasks such as file operations, Git, text processing, and scripts. Commands can also be composed with pipes.
• MCP tends to suit tasks that need extra capabilities or managed services—for example, rendering dynamic webpages, handling OAuth and token refresh, or supporting organizational access controls and audit trails.
• The conclusion is not to choose one exclusively: use CLI when commands map directly to the task, and MCP when its abstraction or governance benefits justify the added tool definitions.

Main source

Martin Keen, presenter for IBM Technology (identified in the video metadata).