Choose a CLI Agent
What are OpenRouter, Claude Code, Codex, and Ollama each?
Pick the right tool and finish a first small task.
📌 Learning goals
- Tell apart LLMs, provider APIs, routers, coding agents, and local runtimes.
- Choose an entry point by the accounts, providers, or local setup you already have — no overall ranking.
- Complete one loop in a demo repo: read → review the plan → confirm → small change → git diff → undo.
Entry conditions
Be able to enter folders and read git status / git diff, with a throwaway demo repo that holds no secrets. If nothing is installed yet, pick one tool from the resource table and follow its official setup.
🧭 Lessons on this site
Read in the suggested order; checkboxes share the same browser progress as the /learn track pages.- 01The AI Coding Tools Map: Recognize the Form Factor First, Memorize Names Later
Cursor, Trae, Claude Code, Codex, OpenClaw, Hermes Agent, DeepSeek Harness, WorkBuddy and Doubao Work sorted into three categories by form factor: IDE-embedded, terminal agent, and orchestration platform. For each category: what it gives you, how much control and visibility it takes away, and how people actually mix them in practice — so you can decide which category your first tool should come from.
13 min - 02What an Agent Harness Is: The Ordinary Program Wrapped Around the Model
The horse-and-harness metaphor makes the concept clear: the layer of engineering code wrapped around the LLM — assembling context, executing tools, running the loop, managing permissions. It explains why the same model can perform worlds apart under a different harness, and why picking an AI tool is, to a large extent, picking a harness rather than a model.
9 min - 03What Is Claude Code: The Reference Template for Terminal Agents
The flow of one real session explains what Claude Code is: a coding agent that lives in your terminal, reading files, editing files, running commands and iterating until the task is done. With installation methods, the CLAUDE.md memory mechanism, form factors beyond the terminal, the billing model, and the three kinds of people who should not start with it.
9 min - 04Installing Claude Code: A Getting-Started Path for the Terminal Agent
Anthropic's official terminal coding agent Claude Code, from system requirements, four install methods (native install script, Homebrew, WinGet, npm), the verification command, and browser sign-in, to permission modes and your first change. Every step spells out the output you should see and the real error messages listed in the official docs.
19 min - 05What Codex Is: OpenAI's Coding Agent, on Two Legs — Local and Cloud
The four forms hiding under the name Codex: the open-source Codex CLI (running in your terminal), the IDE extension, the desktop app, and Codex Web which runs tasks in parallel in the cloud. Covers installation, sign-in and billing models, sandboxing and approval mechanics, and how its form factor differs from Claude Code.
9 min - 06Installing the OpenAI Codex CLI: Sign-In, Permissions, and Your First Task
OpenAI's terminal coding agent Codex CLI, from system requirements, three official installation methods, ChatGPT-account sign-in and API-key sign-in, to the three /permissions presets and non-interactive mode. Every step spells out the output you should see and flags which commands I could not verify against the official documentation.
22 min - 07What OpenClaw Is: An Open-Source Assistant That Runs on Your Own Machine and Lives in Your Chat Apps
OpenClaw is an open-source AI assistant framework: the body runs on your own machine and appears through one Gateway in twenty-odd chat channels at once — WhatsApp, Telegram, Slack, Discord and more. One article covering its three-part architecture, the Skills ecosystem, its "models are swappable" philosophy, and the safety responsibilities you must read before using it.
11 min - 08OpenClaw Installation Guide: One Command to Your Own AI Agent Gateway
Step-by-step installation of the open-source agent framework OpenClaw: the official install script, the first-run setup wizard, installing the Gateway as a background service, opening the web console, and verifying the installation with your first message — plus fixes for the common PATH and npm errors.
16 min - 09The Evolution of the Harness: Six Eras, Each With Its New Capabilities and New Ways to Crash
From pure-prompt chatbots, CoT and ReAct reasoning prompts, and function calling — the real turning point — to MCP standardization, long-running CLI Agents, and always-on multi-Agent orchestration: era by era, what new capability the harness gained each time, and what new failure mode that capability incidentally introduced.
12 min - 10The First Ten Minutes After Installing an Agent Tool: A Seven-Item Verification Checklist
Just installed Cursor, Claude Code, Codex, OpenClaw, Hermes, or any AI Agent tool? Don't assign work yet. Spend ten minutes on seven verifications: it runs, it reaches the model, it reads files, it writes files, permission boundaries, one real task, and where the logs are. Tool-agnostic.
12 min
⚠️ Five guardrails before your first agent
- 1.Never run your first experiment in a directory holding secrets or production permissions.
- 2.Never use the mode that skips all confirmations.
- 3.Never paste API keys, browser tokens, or auth files into a prompt, issue, log, or git.
🎯 Curated resources
| Resource | Who it's for | Priority | Why |
|---|---|---|---|
Official ecosystems anthropics/claude-code | People in the Anthropic ecosystem | ⭐⭐⭐⭐⭐ | Keep the permission prompts; start with a demo repo. |
Official ecosystems openai/codex | People with ChatGPT or an OpenAI API | ⭐⭐⭐⭐⭐ | Confirm approvals, sandbox, and the working directory. |
Official ecosystems anomalyco/opencode | Switching providers or compatible endpoints | ⭐⭐⭐⭐⭐ | AGENTS.md first; also check permission settings. |
Official ecosystems Aider-AI/aider | People who care about git diff and commit flow | ⭐⭐⭐⭐⭐ | Understand its git auto-commit behavior first. |
Official ecosystems NousResearch/hermes-agent | Want an open-source chat-platform assistant | ⭐⭐⭐⭐⭐ | A CLI agent whose gateway reaches many chat platforms. |
Official ecosystems google-gemini/gemini-cli | People with Google auth or Vertex AI | ⭐⭐⭐⭐ | Confirm the auth method and sandbox first. |
Official ecosystems earendil-works/pi | Want a small core plus extensions | ⭐⭐⭐⭐ | The small-core-plus-extensions route. |
Official ecosystems aaif-goose/goose | Want CLI, desktop, and extensions together | ⭐⭐⭐⭐ | Enable only low-privilege extensions first. |
Router / local runtime OpenRouter | Switch providers without switching UIs | ⭐⭐⭐⭐ | A router is not an LLM and does not manage your file permissions. |
Router / local runtime ollama/ollama | Run models locally | ⭐⭐⭐⭐ | A local runtime is not a coding agent; an agent that supports it must call it. |
🛠 Hands-on practice (upstream)
Full exercises & starter codeSummaries from the upstream curriculum; full code, cost, and latency estimates live upstream.
- CLI-1: read-only — have the tool read the demo repo, explain its purpose, find the test command, and propose a small doc-change plan without editing anything.
- CLI-2: confirmed small change — after approving the plan, make one small change visible in git diff and undoable.
- CLI-3: rerun the same request in a second harness — record differences in model/provider, permission prompts, sandbox, and output format instead of picking a winner by feel.
- CLI-4: watch auth fail with fake credentials — distinguish login failure, API-key failure, nonexistent model, and permission blocking, without pasting real secrets into prompts or logs.
✅ Self-check
- I can explain the five roles in my own words, knowing OpenRouter is not an LLM and Ollama is not a coding agent.
- I finished one read-only explanation and plan in a demo repo without handing over secrets.
- I confirmed one small change's diff and can undo it.
- I know my CLI's login method, provider, and approval/sandbox settings.
Adapted from awesome-agentic-ai-zh (MIT, by Wenyu Chiou) v2026.09.23; links checked 2026-08-27. Stars mark learning priority (⭐⭐⭐⭐⭐ = you will get stuck without it), not popularity. MIT License · Curriculum structure last updated 2026-10-03. Content is still being filled in; lessons marked “in progress” are not live yet.