Here’s a pattern worth noticing: your agent can use gh, git, jq, curl, rg, and make — tools already on your machine — without any MCP server, config, or setup.
Simon Willison put it well: “Almost everything I might achieve with an MCP can be handled by a CLI tool instead. LLMs know how to call cli-tool --help, which means you don’t have to spend many tokens describing how to use them — the model can figure it out later when it needs to.”
This matters because of how agents discover tool capabilities. MCP loads all tool descriptions upfront — every turn, every token. CLI tools are lazy-loaded: the agent calls --help only when it actually needs the tool. That’s a significant context window difference when you’re working on a large codebase.
Claude Code leaned into this hard — they deleted their built-in LS tool and replaced it with shell ls. OpenAI Codex shipped with no built-in file tools at all, using sed, cat, and ls for everything.
The practical advice: before building or installing an MCP server, check if a CLI tool already does what you need. Often, it does.
💡 Try this: Ask your agent “list my open PRs” — it will use the gh CLI (you might need to give a hint). No MCP needed.
🔗 Simon Willison on CLI vs MCP
How does your agent interact with external tools?
🟢 Mostly CLI tools — gh, git, jq, etc.
🟡 Mix of CLI and MCP
🔴 Mostly MCP servers
⚪ Haven’t thought about it
Which CLI tools have been most valuable so far?