繁體中文 | 简体中文 | English

CLI Agents Comparison Guide¶

← Back to main path README

📌 This is a reference doc (depth comparison, selection logic, pitfalls, recommended setups). First time touching CLI agents, want step-by-step onboarding → see tracks/cli/A1-cli-intro.en.md (Track A first stop). First want to understand "why does one agent live in a terminal, another in Telegram, another on a Jetson board?" mental model → see resources/agent-paradigms.en.md (5 agent paradigms). Already using one, want to decide / compare / upgrade → stay here.

A cross-branch reference shared by Track A (A1-A3) + all 5 specialized branches: how to choose between Claude Code / Codex / OpenCode / Gemini CLI / goose / Aider / Hermes Agent? Every branch references CLI agents but no single branch "owns" this comparison, so it lives in resources/.

📋 7 Major Terminal CLI Agents¶

Only terminal-based CLI agents are included. IDE-based agents (Cursor / Cline / Continue) live in for-developer. The first 6 numbers verified via gh api on 2026-05-06; Hermes Agent verified on 2026-05-10.

Tool	Provider	License	Primary LLM	Auth / Pricing	Stars
Claude Code	Anthropic (official)	NOASSERTION	Claude	Claude subscription OR Anthropic Console API key	★ 120k+
Codex	OpenAI (official)	Apache-2.0	GPT family	ChatGPT account sign-in OR OpenAI API key	★ 80k+
OpenCode	community (repo now at `anomalyco/opencode`)	MIT	Any (multi-provider)	BYO API key, or built-in OpenCode Zen hosted	★ 155k+
Gemini CLI	Google (official)	Apache-2.0	Gemini	Generous free tier, paid above quota	★ 103k+
goose	Agentic AI Foundation (repo now at `aaif-goose/goose`)	Apache-2.0	15+ providers (incl. Ollama)	BYO API key, or existing Claude / ChatGPT / Gemini subscription via ACP	★ 43k+
Aider	Aider-AI (community)	Apache-2.0	Any	BYO API key	★ 44k+
Hermes Agent	Nous Research	MIT	200+ via OpenRouter / NVIDIA NIM / Zhipu GLM / Kimi / Xiaomi MiMo / MiniMax / HF / OpenAI	BYO API key (multi-provider)	★ 156k+

🎯 Which to Pick? Decide by Use Case¶

Writing papers / literature / research¶

Top pick: Claude Code (long context, strong reasoning, good hallucination resistance). Gemini CLI is the alternative — its million-token context fits whole-PDF / whole-dataset workflows.

Writing code / refactoring a codebase¶

Top pick: Aider (git-native — auto-commits each change, easy to revert) or Claude Code. OpenCode fits when you need to switch between LLMs.

Privacy / offline / no cloud¶

Top pick: goose or OpenCode + local Ollama. Both support BYO LLM and connect to http://localhost:11434/v1 (Ollama default).

Already subscribed to ChatGPT Plus / Pro¶

Top pick: Codex — same account, no separate billing.

Want 1M-token long context + Google ecosystem¶

Top pick: Gemini CLI. Generous free tier and long context are the differentiators. Note: Google service integration (Gmail / Drive / Docs) goes through MCP extensions, not built-in connectors — same setup pattern as other CLIs.

Want to avoid vendor lock-in¶

Top pick: OpenCode > goose > Aider. None tie you to a specific provider; models are swappable.

First time installing a CLI agent — wanting easiest start¶

Top pick: Claude Code. Broad ecosystem, CLAUDE.md mechanism for version-controlled prompts, plenty of community resources when you hit issues.

Want it running on a cloud VM, talking to it via Telegram / Slack / Discord, with mainland China LLMs as primary¶

Top pick: Hermes Agent. Three differentiators: - Decoupled from your laptop — agent runs on a $5 VPS / Modal serverless / Vercel Sandbox; you message it from Telegram / Discord / Slack / WhatsApp / Signal - Model-neutral — supports GLM / Kimi / Xiaomi MiMo / MiniMax, matching the 11 Chinese-ecosystem catalog entries - Built-in self-improving skill loop + cron scheduler — agent autonomously generates skills from interaction, refines them across sessions, runs scheduled jobs unattended - ⚠️ Self-evolving skills is a frontier feature with no independent audit yet; for production tasks, start with low-stakes experiments

📝 Portable Prompts Across CLIs¶

If you want prompts that work across CLI tools (or want to switch without rewriting), follow these principles:

Specify file paths explicitly — "modify src/auth.py" beats "modify that auth file"
Ask for multi-step breakdowns — "first list a plan, then act after I confirm" works in every CLI
Avoid CLI-specific magic — /init /compact are Claude-Code-specific; OpenCode doesn't have them
Use .cursorrules / CLAUDE.md / AGENTS.md for persistent preferences — Claude Code reads CLAUDE.md, Codex reads AGENTS.md, OpenCode reads OPENCODE.md, content can be the same
State review scope clearly — "review only my diff" vs "review the whole repo"

Cross-CLI prompts are usually 5-10% more verbose than CLI-specific ones, but the upside is you can switch tools without rewriting.

⚠️ Common Pitfalls¶

File path handling¶

Windows uses backslashes (C:\Users\...); most CLIs translate internally but sometimes get confused
Recommendation: in git-bash / WSL use forward slashes, avoid weird quoting

Git integration differences¶

Aider auto-commits every change (by design, not a bug)
Claude Code / Codex / OpenCode / goose don't auto-commit by default — manual or via prompt

Default sandbox (each CLI varies; verify against official docs before use)¶

Claude Code: bash writes default to cwd; reads broader (except deny-rule paths)
Codex: in version-controlled folders, Auto (workspace-write + on-request escalation) is recommended; in non-git folders, read-only
goose / OpenCode: relatively permissive — add explicit sandbox / approval rules; don't rely on defaults

Token cost accumulation¶

Running a grep on a large codebase can consume 100k+ tokens
Summarizing a long PDF can hit 500k tokens (Gemini handles this; other tools need to be cost-aware)
Recommendation: estimate cost before each operation; set a monthly cap

Multi-CLI session interference¶

Two CLIs in the same repo (e.g. Claude Code + Aider) can race-condition file edits
Recommendation: one repo, one CLI (unless you genuinely need parallelism)

🔧 Real-World Setups¶

Three common combinations; pick one that fits:

Setup A: Claude Code primary + OpenCode backup¶

Claude Code handles 90% of daily work (code, docs, debug)
OpenCode + Ollama for privacy-sensitive data (medical, financial)
One prompt, runs in either

Setup B: Codex (GPT) + Aider (Claude) mix¶

Codex handles small tasks within ChatGPT Plus quota
Aider with Claude API key handles big refactors (git-native commit convenient)
Separate billing, no interference

Setup C: Gemini CLI primary (long-context scenarios)¶

Whole PDF / whole codebase fed at once
Add Aider for scenarios needing precise git diff
Fits scholars, knowledge workers

Setup D: Hermes Agent + Local Ollama (multi-platform + mainland China LLMs + offline)¶

Hermes Agent runs on a low-cost VPS or your own machine as a multi-platform agent gateway
LLM endpoint can be Ollama (http://localhost:11434/v1), or swapped to providers such as z.ai GLM / Kimi
Chat entrypoint can be Telegram / Slack / Discord; Hermes routes platform messages into the agent workflow
When you want zero Anthropic / OpenAI dependency, this setup fits offline, privacy-sensitive, and low-cost repeat experiments
Step-by-step walkthrough: resources/cookbook.md Recipe 6

Linking Back to Branches¶

Different audiences have different CLI needs:

for-developer: also see IDE-based agents (Cursor, Cline, Continue)
for-everyday-users Tier 2: CLI is the advanced option; try Tier 0 / 1 (Web / Desktop App) first
for-researcher: also see paper-specific tools (paper-qa, gpt-researcher, ChatPaper)
for-knowledge-worker: also see workflow automation (n8n, Make)
for-teacher: CLI is advanced for teachers; start with prompt libraries

Maintenance Notes¶

7 CLI tools' stars / license / pushed_at refreshed quarterly via bash scripts/refresh-stars.py
The CLI market moves fast — new tools require evaluation before inclusion (bar: 30k+ stars, actively maintained, true CLI not IDE)
The comparison table deliberately leaves out "strengths / weaknesses" columns — avoiding subjective bias and letting the use-case section + readers' own judgment do that work