Plain Text Vindicated
Over a decade ago, I read The Art of Unix Programming by Eric S. Raymond. The book champions simplicity and composability: small tools that do one thing well, communicating through plain text rather than complicated binary formats.
That philosophy made Unix the beloved playground of nerds and hackers. Now it's having a moment with AI agents, especially Claude Cowork and Claude Code.
Command-line tools are easy for AI agents to use because invocation is itself plain text. No awkward clicking through a browser or UI, just issuing a command. And if tools work with plain text in both input and output, AI has a much easier time understanding what's going on. Even multi-modal systems handle text better than images.
"Works with plain text" is becoming an important differentiator. Consider where to keep your knowledge base:
A proprietary tool like Notion. Your data lives on their cloud in their database. Exporting is awkward.
A local-first tool like Obsidian. Your data lives on your machine in plain-text markdown.
With Obsidian, Claude can work directly with your notes. Discover them, extract relevant information, and no MCP server required.
Better yet, you only pay for one AI subscription. If I already pay for Gemini, ChatGPT, or Claude as a general-purpose AI tool, I don't want to pay again for AI inside my note-taking, project management, or CRM tools. (And then pay more for the automations connecting them all...)
For tools that must live on the cloud, a strong alternative to MCP servers: provide a capable command-line interface. An AI agent can use the tool directly, and it's useful for humans too.
Claude Code already handles the common Unix tools—find, grep, sed. If you need specific file-manipulation capabilities, wrapping them in a small tool that does one thing well is the Unix way.
The nerds and hackers had it right all along. Now Claude's in on the secret.
