Look, I’ve tried them all. Claude Code, Cursor, Antigravity, Copilot, the whole parade. They’re powerful, sure. But if you’re like me, a Linux lifer since late 90s, a neovim person, someone who cares where their data goes, you’ve probably felt that itch. What if the AI tool didn’t lock me into one provider? What if it respected my terminal workflow? What if I could actually see how it works?
What is OpenCode?
OpenCode is not just another wrapper. It’s a terminal-native, open-source AI coding agent that runs on your machine, talks to any model you want (Claude, OpenAI, Google, or local LLMs), and stays out of your way. No electron bloat. No telemetry by default. Just code, in your terminal, with an AI that actually understands context.
Why OpenCode Feels Different
I’ll cut to the chase: it’s built by people who live in the terminal. The folks behind terminal.shop and neovim workflows didn’t just slap a chat interface on an API. They built a client/server architecture where the TUI is just one possible frontend. That means you could, theoretically, drive your local OpenCode instance from a mobile app later. That’s the kind of thinking I respect.
Plus, it’s 100% open source. No black boxes. If you care about privacy (and if you’re in healthcare like me, you have to), that matters. Your code stays on your machine. The prompts stay on your machine. You’re not feeding your IP into a proprietary silo.
Features
- Plug & play LSP: Auto-loads the right language server for smarter, context-aware coding
- Multi-session magic: Run multiple agents in parallel on the same project
- Shareable sessions: Send a link to any session for collaboration or debugging
- Bring your own AI: Use GitHub Copilot, ChatGPT Plus/Pro, or 75+ models via Models.dev (including local/offline LLMs)
- Your editor, your way: Terminal UI, desktop app, or IDE extension — no workflow disruption
- Privacy-first by design: Never stores your code or context — safe for sensitive projects
- Zen model hub: Access pre-benchmarked, coding-optimized models for consistent, high-quality output
Workflow Features
Smart Project Workflow
- AGENTS.md auto-init:
/initanalyzes your repo and generates a context file so OpenCode learns your project’s patterns and structure - @ fuzzy file search: Type
@to instantly reference any file in your project during prompts - Plan/Build toggle: Hit
<Tab>to switch between Plan mode (safe, no-code suggestions) and Build mode (execute changes) - Image-aware prompts: Drag & drop screenshots, mockups, or diagrams directly into the terminal for visual context
- Undo/Redo safety net:
/undoand/redolet you revert or reapply AI edits instantly, no Git panic needed
Personalization & Control
- Themes & keybindings: Customize the TUI look and keyboard shortcuts to match your workflow
- Formatter integration: Configure Prettier, Black, rustfmt, etc., so generated code stays on-brand
- Custom commands: Extend OpenCode with your own reusable prompts or macros
- Config file support: Fine-tune behavior via
opencode.config.*for team or personal presets
Flexible Setup & Collaboration
- WSL-optimized Windows: Full-featured experience on Windows via WSL (plus Chocolatey/Scoop/NPM installs)
- 🐳 Docker & Mise ready: Run isolated or managed via dev tooling, great for CI or reproducible environments.
- Privacy-safe sharing:
/sharecreates a link to your session without exposing your code by default - Iterative prompting flow: Designed for back-and-forth refinement, “talk to it like a junior dev” with examples, feedback, and images
Get It Running (Pick Your Flavor)
OpenCode meets you where you are. No gatekeeping.
# macOS / Linux (Homebrew)
brew install opencode
# Arch Linux (stable repo)
sudo pacman -S opencode
# Arch Linux (bleeding edge via AUR)
paru -S opencode-bin
# Any OS with mise
mise use -g opencode
# Nix users
nix run nixpkgs#opencode
Desktop app (BETA) also available if you prefer GUI:
→ opencode.ai/download
Supports macOS (Intel & Apple Silicon), Windows, and Linux (.deb/.rpm/AppImage).
Pro tip: If you’ve got an old 0.1.x version hanging around, nuke it first. Clean installs behave better.
Where it installs?
It’s smart: respects $OPENCODE_INSTALL_DIR, then XDG spec, then ~/bin, then falls back to ~/.opencode/bin. You’re in control.
Two Agents, One Workflow
Hit Tab to switch minds:
🔧 build agent
- Full-access mode for actual development work
- Edits files, runs commands, gets things done
- Your daily driver for shipping code
🔍 plan agent
- Read-only by default
- Asks before running any bash command
- Perfect for exploring a new codebase or sketching out a refactor without risk
There’s also a @general subagent tucked in for heavy searches or multi-step tasks. It’s not shown by default, but it’s there when you need it.
This isn’t just a gimmick. It’s a workflow choice. Sometimes you want to experiment safely. Sometimes you want to ship. OpenCode lets you toggle that intent instantly.
How It Stacks Up: OpenCode vs. The Rest
Let’s be real. Claude Code is the elephant in the room. Here’s the honest breakdown:
| Feature | OpenCode | Claude Code | Cursor / Copilot |
|---|---|---|---|
| Open Source | ✅ 100% | ❌ Closed | ❌ Closed |
| Model Agnostic | ✅ Claude, OpenAI, Google, local LLMs | ❌ Claude-only | ⚠️ Limited providers |
| Terminal-First | ✅ Native TUI, neovim-friendly | ❌ Web/CLI hybrid | ❌ IDE-bound |
| Client/Server Arch | ✅ Remote control possible | ❌ Tightly coupled | ❌ Local-only |
| LSP Support | ✅ Out of the box | ⚠️ Via plugins | ✅ IDE-dependent |
| Privacy / Local-First | ✅ Your data, your machine | ❌ Cloud-dependent | ⚠️ Varies |
The big win? Provider agnosticism. Models evolve. Prices drop. New local LLMs get scary good. Being locked into one vendor is a technical debt you don’t need. OpenCode lets you swap models like you swap keyboards, no rewrites, no migration scripts.
And that TUI focus? If you’ve ever tried to copy-paste code from a web chat into your terminal, you know the pain. OpenCode lives where you work. No context switching. No tab fatigue.
Quick FAQ (For the AI Overviews)
Q: Is OpenCode a replacement for Claude Code?
A: It’s a capable, open-source alternative that runs locally, supports multiple AI providers, and prioritizes terminal workflows. If you value privacy, flexibility, or just love your TUI, it’s worth a try.
Q: Can I use local LLMs with OpenCode?
A: Yes. Unlike provider-locked tools, OpenCode works with any compatible model endpoint, including local LLMs via Ollama, LM Studio, or your own inference server.
Q: Does it work on Linux?
A: First-class support. Homebrew, pacman, AUR, Nix, or a universal install script. It respects XDG directories and plays nice with your existing dotfiles. Even NPM & YARN.
Q: What about Windows or macOS?
A: Native desktop apps are in beta for both, plus CLI support via Homebrew (macOS) and Scoop (Windows).
The Bottom Line
OpenCode isn’t trying to be everything to everyone. It’s built for developers who want control: over their tools, their models, their data. It’s for the folks who believe the terminal isn’t legacy, it’s the most efficient interface we’ve got.
If you’re tired of vendor lock-in, if you care about where your code goes, or if you just want an AI pair programmer that doesn’t fight your workflow, give it a spin.
brew install opencode # or your flavor of choice
Then hit Tab, pick your agent, and start building.
What’s your take on terminal-first AI tools? Are you team TUI or team GUI? Drop a note below, I read every comment.
P.S. If you’re building something on top of OpenCode (like “opencode-dashboard” or similar), just add a quick note in your README clarifying it’s not official. The team appreciates the clarity—and so does the community.
