how to choose a CLI coding agent

start from the subscription you already pay for

Most CLI coding agents come bundled with a chat subscription. Claude Code is included in Claude Pro and Max, Codex is included in ChatGPT plans down to the free tier, GitHub Copilot CLI ships with every Copilot plan, and Kiro has a free tier with 50 credits a month. If you already pay one of these vendors, their agent is the cheapest place to start.

open source and local models

Codex, opencode, goose, Aider, and Kimi Code publish their source. For models, opencode, goose, and Aider connect to almost any provider, local models through Ollama included, and Codex runs local open-source models through Ollama or LM Studio. If your code can't leave your hardware, these are the candidates.

one model family or a switchboard

Claude Code and Muse Code are built around their vendor's own models. GitHub Copilot CLI, Kiro, and Antigravity CLI take the opposite bet: a model picker spanning Claude, GPT, Gemini, and open-weight models, so one subscription covers several families.

which agent for which job in crystl

Every agent here runs in crystl. These groupings are claims about crystl fit, not model quality: which host surfaces are wired and tested for each agent.

running the lead shard

The agent that owns a long task needs the full surface: approval cards, ask-user questions, transcript history, and resume after a restart.

• Claude Code the reference integration — every surface tested
• Codex full parity on cards, transcripts, and resume

fanning out workers

Workers run gated: tool calls raise approval cards, and the lead waits on their turn reports instead of polling screens.

• Claude Code cards and turn ends fully wired
• Codex cards and turn ends fully wired
• opencode cards via an injected plugin; a missed answer falls back to its own prompt
• goose cards in its default auto mode only
• Kimi Code cards can deny but not pre-approve
• GitHub Copilot CLI cards for the tools it gates itself; an allow-all session drops them

quests and multi-agent chat

A quest party needs transcripts crystl can read back and prompts it can deliver mid-session.

• Claude Code
• Codex
• Antigravity CLI transcripts read back; other surfaces unverified
• Kimi Code quest prompts arrive as follow-up messages

open and local models

Provider-agnostic agents point at whatever endpoint you configure, so open models can take the jobs they are good at.

• opencode open source, bring any provider
• goose open source, bring any provider
• Aider open source, provider-prefixed model ids, local models included
• Claude Code runs against a configured model endpoint too
• Codex runs against a configured model endpoint too

Compatibility is descriptive

This catalog explains the host surfaces crystl has tested. Your local Crystl installation remains the source of truth for what runs, which hooks are installed, and how each agent is launched.