OpenCode

If you write code with an AI agent in your terminal, you have probably already switched to OpenCode, or you are about to. The MIT-licensed coding agent crossed 179,000 GitHub stars and around 7.5 million monthly developers by late June 2026, ships a new release almost every day, and works with 75-plus model providers instead of locking you to one lab. Then on June 18, Google pulled the plug on Gemini CLI for free and consumer users, and the migration turned into a stampede. Here is what OpenCode is, why it won, and how to point it at the open models we have been covering.

What OpenCode is

OpenCode is an open-source AI coding agent that runs in your terminal, plus a beta desktop app and a VS Code extension. It is not a chat box. It reads your files, edits code, runs shell commands, and tests its own work, the full agentic loop. It is built by Anomaly, the team formerly known as SST. (One note to avoid confusion: the original Go version split off and continues as "Crush"; the TypeScript OpenCode at github.com/anomalyco/opencode is the one everyone means now, and the old sst/opencode URL just redirects there.)

The defining choice is that OpenCode ships no model of its own. It is model-agnostic, talking to 75-plus providers through Models.dev: Claude, GPT, Gemini, DeepSeek, and any OpenAI-compatible endpoint, including local servers. It supports MCP (the Model Context Protocol, the now-standard way agents call external tools) and LSP (Language Server Protocol) for real code intelligence. It ships built-in agents (a read-write "build" agent, a read-only "plan" agent, a "general" subagent) and lets you define your own.

Why it won: the Gemini CLI exodus

OpenCode was already the most-starred open coding harness. Then Google handed it the market. On June 18, 2026, Google shut down Gemini CLI for free, Pro, and Ultra consumer users with no grace period, and is ending consumer Code Assist for GitHub on July 17. The replacement, Antigravity CLI, is closed-source and cut the free tier from roughly 1,000 requests a day to about 20, a 98 percent reduction.

For a developer who built a workflow around a free, open terminal agent, that is a forced migration. OpenCode is the obvious landing spot, because it still talks to Gemini models through the API if you want them, alongside 74 other providers, with no vendor able to switch it off. Qwen Code, Alibaba's Apache-2.0 fork of the old Gemini CLI, picked up the overflow. The lesson writes itself: an open, model-agnostic tool cannot be discontinued out from under you.

One honest caveat about benchmarks

You will see people quote SWE-Bench or Terminal-Bench numbers "for OpenCode." Ignore them. Because OpenCode has no built-in model, it has no intrinsic score; its ceiling is whatever model you feed it. Point it at Claude Opus 4.8 or GPT-5.5 and you get top-of-leaderboard coding (Opus 4.8 lands around 74 percent on Terminal-Bench 2.1 and 88 percent on SWE-Bench Verified). Point it at a weak local model and you get weak results. The tool is the harness, not the brain.

Install it and point it at an open model in 10 minutes

Install is one line, by whatever package manager you already use.

# pick one
npm i -g opencode-ai@latest
brew install anomalyco/tap/opencode
curl -fsSL https://opencode.ai/install | bash

# then just run it inside a project
opencode

The interesting part for this site is wiring it to an open model instead of a closed API. OpenCode reads an opencode.json in your project. Here it is pointed at a local server (Ollama, llama.cpp, or LM Studio) over the OpenAI-compatible API; the same pattern works for a rented GPU running GLM-5.2 or Qwen3.6:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "local": {
      "npm": "@ai-sdk/openai-compatible",
      "options": { "baseURL": "http://localhost:11434/v1" },
      "models": { "qwen3.6:27b": { "name": "Qwen3.6 27B (local)" } }
    }
  },
  "model": "local/qwen3.6:27b"
}

Switch models mid-session with the /models command. That is the whole pitch: one agent, every model, including the open ones you can actually own. Pair it with GLM-5.2 for cheap agentic coding, or a local Qwen when you want it air-gapped.

Who should use it

Anyone who wants a serious terminal coding agent without lock-in: Claude Code users who want model freedom, ex-Gemini-CLI users who just got evicted, and anyone running local or open-weight models who needs a harness that speaks their endpoint. If you are deep in a single closed ecosystem and happy there, you may not need it. Everyone else: the switching cost is one install command.

Sources and further reading

• OpenCode official site
• OpenCode on GitHub (MIT)
• OpenCode docs: providers and local models
• 9to5Google: Gemini CLI shutting down

Tested on: not benchmarked. OpenCode is model-agnostic and has no intrinsic score; its quality equals the model you connect. The facts here (stars, version v1.17.11, 75-plus providers, the June 18 Gemini CLI shutdown) are verified against the GitHub repo and primary coverage.
Date checked: 2026-06-26