Tell an AI agent "rename 400 product photos to SKU format, resize to 1200x1200, and sort by category" — done in seconds. Would take a human 2 hours. How does the agent do it? The answer is CLI.
What Is This?
Ben Tossell (Ben's Bites) summed up how AI agents work in one line: "Agents are LLMs with tool-use. They don't just respond — they can go and do things for you". And the most fundamental form of those tools is CLI (Command Line Interface).
CLI is a text-based way to control software. Type a command, something happens. Since agents communicate in text, CLI is a natural fit.
The most basic is bash — the general-purpose command line on your computer. ls lists files, mkdir creates folders, mv moves files, mogrify resizes images. Agents combine these to automate complex tasks.
But bash is just the start. There are purpose-built CLIs for specific jobs:
| CLI Tool | Purpose | What Agents Can Do |
|---|---|---|
| bash | General system commands | File management, scripts, data processing |
| Stripe CLI | Payments/subscriptions | Pull revenue data, manage subs, test payments |
| Playwright | Browser automation | Navigate, click, fill forms, screenshots |
| AWS CLI | Cloud infrastructure | Spin up servers, manage DBs, scale |
| Vercel CLI | Web deployment | Deploy a site live in one command |
The more CLIs you give an agent, the more it can do. Add Stripe CLI and it pulls revenue numbers. Add Playwright and it browses the web. Add Vercel and it deploys what it builds. Your job is to make sure it has the right tools for the task.
What Changes?
As of 2026, there are over 15 terminal-based AI coding agents. A year ago it was just "Copilot vs Cursor" — now it's a completely different landscape.
| IDE-based (Cursor etc.) | CLI-based Agents | |
|---|---|---|
| Workflow | Code suggestions, autocomplete | Plan → execute → verify autonomously |
| Scope | Code in editor only | Files, terminal, git, external APIs |
| Autonomy | Human approves | Agent iterates on its own |
| Extensibility | Plugin ecosystem | Any CLI can be added as a tool |
According to Tembo's analysis, CLI coding agents fall into three groups:
Aider leads with 39K+ GitHub stars and 15 billion tokens processed weekly, while OpenCode is surging with 95K+ stars. Gemini CLI has the lowest barrier with its free tier (60 requests/min, 1,000/day).
Getting Started
- Choose based on your model
Anthropic → Claude Code, OpenAI → Codex, Google → Gemini CLI. If committed to one model, start with the native tool. - Free start? Go Gemini CLI
1,000 free requests/day with just a Google account. 1M token context window is great for large codebases. - Need model flexibility? Aider or OpenCode
Support almost every LLM, fully open source. You only pay model costs directly. - Extend with purpose-built CLIs
Add Stripe CLI, Playwright, Vercel CLI to your coding agent — capabilities multiply exponentially.
Quick Selection Guide
Want high autonomy → Claude Code, Droid, Warp
Need human approval → Cline (VS Code), Kiro (spec-driven)
Zero budget → Gemini CLI, Aider, Goose, OpenCode
