opencode Cheat Sheet — the open terminal agent

The philosophy is in the name: open. opencode gives you the full terminal-agent experience — slick TUI, tools, sessions — and you pick the model underneath. Claude, GPT, Gemini, a local model on your box… all work. If vendor lock-in gives you allergies, this is your playground.

Install & setup

Command	What it does
npm install -g opencode-ai	Install (or curl -fsSL https://opencode.ai/install | bash)
opencode auth login	Connect a provider — dozens supported
opencode	Open the TUI in the current directory
opencode run "your prompt"	Non-interactive mode for scripts
opencode upgrade	Update

Inside the TUI

Command	What it does
/help	All commands
/init	Generate AGENTS.md — the project memory file
/models	Pick a model from any connected provider
/sessions	Switch between sessions
/share	Share the conversation via link
/undo	Revert the last edit — your safety valve
/compact	Summarize conversation to free context
/exit	Quit

The nice move: modes

opencode ships with modes you cycle via Tab: build mode (normal work — edit & run) and plan mode (read & plan only — touches nothing). You can define custom modes with their own prompts and tools in opencode.json.

The verdict

Its killer use: experimentation and comparison. One task — how does Claude solve it vs GPT vs a cheap open-weight model? Same harness, swap with /models, compare with your own eyes. There's no cleaner way to understand real model differences at work — beyond benchmark tables.