OpenCode AI: install and connect multiple LLM providers

What OpenCode is

OpenCode is a terminal-based coding agent that reads your project, plans changes, and executes them. It is comparable to Claude Code or Codex CLI but with one defining trait: it works with any LLM provider. You pick the model, OpenCode handles the agent loop (reading files, planning, writing code, running commands).

The project is written in Go, which means fast startup times and a single binary with no runtime dependencies. No Node.js, no Python, no Docker. Download the binary and run it.

On Terminal-Bench (a benchmark for terminal-based coding agents), OpenCode scored 51.7% with Claude Opus 4.5 as the backend model. That is competitive with other open-source agents, though purpose-built tools like Claude Code score higher with the same model due to tighter integration.

Install OpenCode

There are three ways to install:

Option A: Go install (if you have Go 1.22+)

go install github.com/opencode-ai/opencode@latest

Option B: Homebrew (macOS)

brew install opencode-ai/tap/opencode

Option C: Download binary

Download the latest release from the GitHub releases page for your platform. Extract and move the binary to a directory in your PATH:

tar -xzf opencode_*.tar.gz
mv opencode ~/.local/bin/

Verify the install:

opencode --version

Configure your first provider

Initialize the config file:

opencode init

This creates ~/.config/opencode/config.json. Open it and set your provider.

Anthropic (Claude)

{
  "provider": "anthropic",
  "model": "claude-sonnet-4-20250514",
  "api_key": "sk-ant-your-key-here"
}

OpenAI

{
  "provider": "openai",
  "model": "gpt-4o",
  "api_key": "sk-your-key-here"
}

xAI (Grok)

{
  "provider": "xai",
  "model": "grok-4",
  "api_key": "xai-your-key-here"
}

Local Ollama

{
  "provider": "ollama",
  "model": "llama3.1:70b",
  "base_url": "http://localhost:11434"
}

Make sure Ollama is running (ollama serve) and the model is pulled (ollama pull llama3.1:70b) before starting OpenCode with this config.

Set up shell aliases for fast switching

Rather than editing the config file every time you want a different model, set up shell aliases that override the provider on launch. Add these to your ~/.zshrc or ~/.bashrc:

# OpenCode aliases
alias oc="opencode"
alias oc-claude="OPENCODE_PROVIDER=anthropic OPENCODE_MODEL=claude-sonnet-4-20250514 opencode"
alias oc-gpt="OPENCODE_PROVIDER=openai OPENCODE_MODEL=gpt-4o opencode"
alias oc-grok="OPENCODE_PROVIDER=xai OPENCODE_MODEL=grok-4 opencode"
alias oc-local="OPENCODE_PROVIDER=ollama OPENCODE_MODEL=llama3.1:70b opencode"

Reload your shell:

source ~/.zshrc

Now you can switch providers with a single command. Type oc-gpt for OpenAI, oc-claude for Anthropic, or oc-local for a local model. Each alias sets the provider and model via environment variables, overriding the config file for that session.

Your first coding session

Navigate to a project and start OpenCode:

cd ~/projects/my-app
oc

Type a task:

Add a health check endpoint at GET /health that returns the app version and uptime

OpenCode reads your project structure, identifies the framework and routing pattern, generates the endpoint code, and shows you the proposed changes. Review the diff and approve. The workflow is similar to Claude Code or Codex CLI: read project, plan, generate, review, apply.

When to use which provider

Not every model is equally good at every task. After testing OpenCode with multiple providers, here is a rough guide:

• Anthropic Claude Sonnet: Best for multi-file refactoring, code review, and tasks that require understanding relationships between files. Strong at following existing code patterns.
• OpenAI GPT-4o: Good general-purpose option. Performs well on standard coding tasks and has broad language support. Sometimes generates more verbose output than Claude.
• xAI Grok: Strong at reasoning-heavy tasks and technical explanations. Newer to the coding agent space, so ecosystem support is still catching up.
• Local Ollama (Llama 3.1 70B): Free and private. Acceptable for simple edits, boilerplate, and tasks where you do not want code leaving your machine. Noticeably worse than cloud models on complex multi-file work.

The alias approach lets you pick the right tool for each task without friction. Start with oc-claude for complex architecture work, drop to oc-local for quick private edits.

Troubleshooting

• opencode: command not found after go install: Ensure $GOPATH/bin (typically ~/go/bin) is in your PATH.
• API errors (401): Double-check the API key format for your provider. Anthropic keys start with sk-ant-, OpenAI with sk-, xAI with xai-.
• Ollama connection refused: Start the Ollama server with ollama serve before launching OpenCode. Confirm the model is pulled with ollama list.
• Slow responses with local models: Models above 30B parameters need significant RAM and ideally a GPU. On CPU-only machines, expect 30 to 60 second response times for complex tasks.

Next steps

With OpenCode configured, try setting up project-specific configs (drop a .opencode.json in any project root to override global settings), explore the built-in tools for shell execution and web browsing, and benchmark different models on your specific codebase. For a comparison of all terminal agents, see the agentic AI overview.

Frequently asked questions