OpenCode — The Open-Source AI Coding Agent Framework

If you’ve used Claude Code or Cursor, you know how powerful an AI coding agent can be. But what if you want something open-source, provider-agnostic, and terminal-native — something that works with any LLM, in any environment, without being locked into a single ecosystem?

That’s exactly what OpenCode delivers.

What is OpenCode?#

OpenCode is an open-source AI coding assistant and agent framework that runs directly in your terminal. It started as a Go-based CLI and has evolved into a dual-layer architecture:

• Backend: A high-performance HTTP server (Bun + Hono + TypeScript)
• Frontend: A rich Terminal UI (Go + Bubble Tea)

OpenCode gives AI agents fine-grained access to your files, terminal shell, Language Server Protocol (LSP) diagnostics, and external services via the Model Context Protocol (MCP). It’s designed to be a fully open-source alternative to proprietary tools like Claude Code — with the freedom to use any model provider you choose.

Why OpenCode?#

1. Provider-Agnostic — 75+ LLM Providers#

Unlike Claude Code (tied to Anthropic) or GitHub Copilot (tied to OpenAI/Microsoft), OpenCode works with 75+ LLM providers through Vercel’s AI SDK and Models.dev. Use any model you want:

• Anthropic: Claude 3.5 Sonnet, Claude 4 Opus
• OpenAI: GPT-4o, o3, GPT-5
• Google: Gemini 2.5 Pro, Gemini 2.0 Flash
• Open-Source: Llama 3, DeepSeek-V3, Qwen 3
• Local: Ollama, LM Studio, vLLM
• Cloud: AWS Bedrock, Azure, Groq

You can even configure different models for different agents — use a cheap, fast model for planning and a powerful reasoning model for building.

2. Terminal-Native with a Rich TUI#

OpenCode runs in your terminal with a full-screen TUI built on Bubble Tea. It’s designed for keyboard-driven workflows:

3. Advanced Agent Architecture#

OpenCode uses a primary agent + subagent hierarchy:

Primary agents:

• Build: Full tool access — can read, write, edit files, run shell commands, access LSP diagnostics. This is your implementation agent.
• Plan: Read-only with ask-only permissions. Analyzes code and formulates strategies without changing anything.

Subagents (summonable via @):

• General: Multi-step task execution
• Explore: Fast codebase navigation
• Scout: External dependency research and documentation lookup

4. Model Context Protocol (MCP) Integration#

OpenCode acts as an MCP client, dynamically loading tools from external MCP servers — both Stdio-based and SSE-based. Every MCP tool inherits the same permission system as native tools, so you control exactly what the AI can access.

5. Language Server Protocol (LSP) Support#

OpenCode connects to LSP servers (like gopls, typescript-language-server, rust-analyzer) to feed real-time diagnostics to the LLM. The AI sees errors, warnings, and hints as it edits code — just like an IDE would.

6. Granular Permission Controls#

Security is built in at every level. OpenCode supports three permission modes:

• Allow: Auto-approve (safe for read-only commands like git status)
• Ask: Prompt for approval (default for write commands)
• Deny: Always block (dangerous operations)

Permissions can be set globally, per-tool, per-file, or per-command. This means you can auto-allow git status while requiring approval for git push — all configured in opencode.json.

7. Auto-Compact Context Management#

OpenCode automatically monitors token usage. When you approach 95% of the model’s context window, it summarizes the conversation history to avoid “out of context” errors. This means you can have long, productive sessions without hitting model limits.

8. Plugin System & Customization#

Custom Commands: Create Markdown files in ~/.config/opencode/commands/ or .opencode/commands/ to define reusable multi-step workflows with named parameters.

Custom Agents: Configure specialized agents with custom system prompts, temperature, step limits, and permission boundaries — all via opencode.json or Markdown files.

Middleware Hooks: Plugins can hook into tool execution lifecycles to audit logs, manage telemetry, or inject custom verification.

Plugin Ecosystem: Tools like opencode-scheduler ↗ extend OpenCode with scheduling, background execution, and more.

Installation#

OpenCode installs on all major platforms:

# Official install script
curl -fsSL https://opencode.ai/install | bash

# Or via npm
npm install -g opencode-ai

# Or via Homebrew (macOS/Linux)
brew install anomalyco/tap/opencode

# Or via Bun
bun install -g opencode-ai

Getting Started#

# Navigate to your project
cd /path/to/project

# Launch OpenCode
opencode

# Inside the TUI:
# 1. Run /connect to add your API provider
# 2. Run /init to generate AGENTS.md with project context
# 3. Start coding!

OpenCode also supports non-interactive scripting mode for CI/CD:

# Run a single prompt and get JSON output
opencode -p "Find all TODO comments in this project" --json

How It Works#

OpenCode runs a client/server architecture:

1. The Go TUI (client) connects to a Bun/Hono HTTP server (backend)
2. The server orchestrates LLM calls through the AI SDK, managing tool calls, context windows, and agent hierarchy
3. Client SDK code is generated via Stainless from an OpenAPI spec — any app (mobile, web, script) can interface with OpenCode programmatically
4. The model is the brain — native tools (file read/write, bash, grep, glob, LSP diagnostics, webfetch) are its hands

OpenCode vs. Alternatives#

The Ecosystem Grows#

OpenCode’s plugin architecture has spawned a growing ecosystem. The opencode-scheduler ↗ plugin, for example, lets you run AI agents on a schedule using your OS native scheduler — perfect for autonomous deal hunting, weekly code audits, or uptime monitoring.

With its open-source MIT license, provider flexibility, and powerful plugin system, OpenCode is becoming the go-to choice for developers who want an AI coding agent that works their way.

References#

• OpenCode GitHub Repository ↗
• OpenCode Official Website ↗
• OpenCode Scheduler Plugin ↗