home / why

Why CLIs matter in the age of agents

And why the biggest companies in the world are shipping them right now.

The shift that's happening

Every major company is shipping a CLI: Google released gws for Workspace, GitHub has gh, Vercel has v0, Cloudflare has wrangler, OpenSea just launched opensea-cli, and shadcn shipped v4 today.

This isn't a coincidence. CLIs are the interface layer for the agent era.

AI agents don't click buttons. They don't navigate web UIs. They run commands, parse output, and compose tools together. The terminal is the most natural interface an agent can use — and it's been battle-tested for 50 years.

CLIs vs MCP servers: the 275x difference

Connect a GitHub MCP server and 93 tools get dumped into your context window — 55,000 tokens gone. Before you've even asked a question. Stack GitHub + Jira + a database + Microsoft Graph? 150,000+ tokens. Just for plumbing.

The same task via gh issue list --json? ~200 tokens.

That's not a minor difference. That's a 275x difference.

CLIs

  • LLMs already know them — trained on millions of man pages
  • Unix pipes: 50+ years of composability
  • Auth already solved (gh auth, aws sso)
  • Instant debugging — no two-process mystery
  • Minimal tokens: ~200 per call
  • Zero dependencies — just a binary

MCP Servers

  • Enterprise auth: OAuth, audit trails, scoped permissions
  • Multi-tenant SaaS with fine-grained access
  • Marketplace discovery (emerging)
  • Context pollution: 55k+ tokens per server
  • Fragile: stdio process management, version conflicts
  • Hard to debug — two-process architecture

"CLIs are super exciting precisely because they are a legacy technology, which means AI agents can natively and easily use them."

— Andrej Karpathy

It's not CLI vs MCP — it's knowing when to use which

The debate isn't binary. The real answer:

CLI for dev workflows

Personal automation, coding agents, shell pipelines, local development. Fast, lightweight, composable. Your agent runs gh pr list --json and gets exactly what it needs.

MCP for enterprise governance

Multi-tenant SaaS, audit trails, OAuth scopes, managed deployments. When you need fine-grained access control and can't give agents shell access.

Skills for the best of both worlds

Agent Skills (agentskills.io) are markdown files that teach agents how to use CLIs. Zero overhead, human-readable, version-controlled. The glue layer.

The "CLI Army" pattern

Peter Steinberger pioneered this approach while building OpenClaw — the fastest-growing open-source project on GitHub. Instead of wiring up MCP servers for every service, he built a dedicated CLI for each service:

gog
Google Workspace
wacli
WhatsApp
imsg
iMessage
spogo
Spotify
sonoscli
Sonos
remindctl
Apple Reminders
summarize
URL/YouTube
sag
ElevenLabs TTS
goplaces
Google Places
ordercli
Food Delivery
eightctl
Eight Sleep
metcli
Meta/Facebook

Each tool is small (most under 2,000 lines of Go), wraps a single API, outputs JSON, and composes with everything else. Together, they give an AI agent superhuman access to the digital world — without a single MCP server.

The pattern: one CLI per service, structured output, composable by default.

Explore steipete's repos →

Recent CLI launches

Companies are shipping new CLIs every week. Here are the latest:

Companies betting on CLIs

These aren't hobby projects. These are official, company-backed CLI tools.

What this means for you

If your product has an API but no CLI, you're leaving agent integration on the table. The barrier to giving AI agents access to your service is a single Go binary that wraps your REST API and outputs JSON.

The companies that ship CLIs now will be the ones AI agents recommend, use, and build around. The rest will need an MCP adapter — an extra layer of complexity that agents would rather avoid.

The terminal isn't going away. It's becoming the most important interface of the decade.