A1 — CLI Agent Intro & Selection¶

繁體中文 | 简体中文 | English

← Back to main path README · Track A: CLI Power User — Stop 1

⏱ Time estimate: 1 week (~5-10 hours)

📋 Chapter structure: Learning goals → Entry conditions → Required reading → Hands-on exercises → Curated Projects → Self-check 🔑 Key term: this page only uses CLI agent (an AI tool that runs in the terminal). MCP / Skill / plugin and other ecosystem terms are introduced where they first appear in A2 / A3. Full glossary: resources/glossary.en.md.

After Stages 0-2, you want to use existing CLI agents to get real work done — not write agent code yourself, just use existing tools to complete tasks first. This track is for you. First stop: pick a CLI agent and get it running.

📌 Learning Goals¶

Know the differences between 7 mainstream CLI agents (Claude Code / Codex / OpenCode / Gemini CLI / goose / Aider / Hermes Agent)
Pick a first CLI tool based on your scenario
Complete install + auth + your first real task (not a hello world)
Know when to switch / add a second CLI

🚪 Entry Conditions¶

You should already: - Have completed Stage 0's Exercise: CLI (basic command-line literacy) - Have a Claude / OpenAI / Google account (paid not required) - Be comfortable with prompt design (Stage 2)

📚 Required Reading¶

resources/agent-paradigms.en.md ⭐ — the 5-paradigm map of the agent landscape; read this first to see where CLI agents sit (Type 2 + Type 3) in the wider ecosystem
resources/cli-agents-guide.en.md ⭐ — the core reference for this track. 7 mainstream CLI agents side by side, use-case picks, real-world setups
Anthropic — Claude Code Quickstart — official install
OpenAI — Codex Quickstart — Codex install + auth

🛠 Hands-on Exercises (foundational, illustrative)¶

Exercise CLI-1: Install + first run¶

Finish it in 3 steps:

Install: follow your chosen CLI's quickstart (each CLI's official docs should have a ≤5-minute install guide)
Pick a low-risk real task: don't write "hello world" — choose something you were already going to do today, e.g. "organize my Downloads folder and move all PDFs to ~/Documents/PDFs"
Observe 3 things: how it decomposes the task, when it asks for confirmation, and what output format it uses

→ Real tasks are what make the difference between an agent and a chatbot visible.

Exercise CLI-2: CLI's built-in system prompt file¶

Claude Code → write a CLAUDE.md at the repo root
Codex → write AGENTS.md
Gemini CLI → write GEMINI.md
goose / OpenCode → see each tool's docs

Put 3 things in it: "your persona / preferred code style / things you can't do". Then run a task and observe behavioral differences.

Exercise CLI-3: Run a second CLI alongside¶

Install a second CLI (suggest Codex or OpenCode as backup). Run the same prompt and compare output style, speed, cost. Not to pick a winner — to learn that "different CLIs solve the same problem from different angles".

Exercise CLI-4: Auth corner cases¶

Deliberately break your API key (one wrong character) and see how the CLI errors out. Then "correct key but wrong model name". Production usage will hit auth issues — step on these now.

🎯 Curated Projects¶

7 Mainstream CLI Agents¶

Detailed comparison (stars, license, strengths, recommended use cases) in resources/cli-agents-guide.en.md. Quick entry points here:

anthropics/claude-code ⭐⭐⭐⭐⭐¶

★ 120k+ — Recommended first CLI agent. Built-in SKILL / plugin ecosystem, CLAUDE.md prompt system, most complete community resources.

openai/codex ⭐⭐⭐⭐⭐¶

★ 80k+ — Top pick if you already subscribe to ChatGPT Plus / Pro; same account works in the terminal.

sst/opencode ⭐⭐⭐⭐⭐¶

★ 155k+ — Open-source, not tied to any LLM provider, fastest community iteration. Pick this for self-hosting or no vendor lock-in.

google-gemini/gemini-cli ⭐⭐⭐⭐¶

★ 103k+ — When you want 1M-token long context for big codebases / large PDFs.

block/goose ⭐⭐⭐⭐¶

★ 43k+ — 15+ provider support (incl. Ollama); use existing Claude / ChatGPT / Gemini subscriptions. Now at aaif-goose/goose (AAIF / Linux Foundation).

Aider-AI/aider ⭐⭐⭐⭐⭐¶

★ 44k+ — git-native, auto commit / branch. Pick this when you want clean git workflow with code edits.

NousResearch/hermes-agent ⭐⭐⭐⭐⭐¶

★ 142k+ — Nous Research's auto-evolving agent. Three differentiators: (1) the agent runs on a cloud VM and you message it from Telegram / Discord / Slack / WhatsApp / Signal; (2) model-neutral — supports GLM / Kimi / Xiaomi MiMo / MiniMax and other Chinese-ecosystem LLMs; (3) built-in cron scheduler + autonomous skill-evolution loop (★ data as of 2026-05; check the official GitHub for current numbers). ⚠️ Auto-evolving skills are experimental, lack third-party independent audits, and should be safety- and maintenance-verified before production use; start in low-risk contexts.

Adjacent tools¶

LM Studio ¶

Closed-source desktop app — drag-and-drop UI for local LLMs. Try this first if you're on Windows / Mac and want local LLM without command line.

Ollama ¶

★ 170k+ — Local LLM runner; pairs well with OpenCode / goose (and any tool with OpenAI-compatible base_url). See Stage 1 — Local LLM section.

✅ Self-Check Before A2¶

Can you: - [ ] Articulate the core differences between the 7 mainstream CLIs (3-4 without checking the table) - [ ] Have a working primary CLI (installed, authed, ran 5+ real tasks) - [ ] Written your own CLAUDE.md / AGENTS.md / GEMINI.md - [ ] Run a second CLI at least once, know the style differences

If yes → proceed to A2 — CLI Workflow Patterns.

If no → don't skip. Sloppy CLI usage isn't productive CLI usage; do Exercises CLI-1/2 at least 3 more times.

💡 Reminder for Track A learners¶

A CLI agent is not "the same thing with a different UI" as Claude.ai / ChatGPT web — it can read/write files on your machine, run shell commands, modify git. This capability difference deserves caution before use: - Week 1: review the plan before letting it execute (or use --dry-run) - Don't let CLI commit directly to production codebases yet - Put sensitive data (keys, contracts, medical records) in .cursorignore / .claudeignore to exclude