🔌 API + CLI + Skills = the Future of AI Agents

DHH's recent post offers a glimpse into the future:

"So while we keep cooking on actually-useful native AI features in Basecamp, we're launching a fully agent-accessible version today. We've revamped our API, created a brand-new CLI, and wrapped it all in a skill to teach agents how best to use it all. It works remarkably well, and it's really fast too."

Especially check out the skill.

This matches what Peter Steinberger, the author of OpenClaw, has been saying:

"You know, you know what I love? That half a year ago, like everyone was talking about MCPs … and I was like, “Sc*** MCPs. Every MCP would be better as a CLI." And now this stuff doesn’t even have MCP support. I mean, it, it has with asterisks, but not in the core lair, and nobody’s complaining."

Why do CLIs work so well?

• No MCP server needed.
• Atomic process with an exit code and text stdout/stderr — exactly the kind of deterministic loop agents need for self-correction.
• Compose naturally via pipes, redirection, and shell scripting.
• Low integration overhead.
• Deploy anywhere.
• An agent can figure out how to use a CLI by running <cli command> --help.
• CLIs mix API calls with local operations.

With an existing CLI, creating a Claude skill for it is straightforward.