OpenCode: The Open-Source AI Coding Agent With 216,000 GitHub Stars

OpenCode calls itself "the open source AI coding agent," and the numbers back up the ambition. The repo at anomalyco/opencode has over 216,000 GitHub stars, nearly 29,000 forks, and an MIT license, which means you can use it, fork it, and build on it without paying anyone anything. Its makers claim 16 million developers use it every month.

The pitch is simple: everything the paid coding agents do, in an interface you already live in. It edits files, runs shell commands, searches your codebase, and reasons through multi-step tasks. But unlike the closed alternatives, you pick the model, you pick the surface (terminal, desktop app, or editor), and you can see exactly what it is doing because the whole thing is open source.

If you have been paying for an AI coding subscription and wondering whether the free tier of the open-source world has caught up, this is the repo everyone points to first.

OpenCode ships with two agents and you flip between them with the Tab key.

build is the default. It has full access: read, write, edit files, run bash commands, the works. This is the agent you leave running when you want work done.

plan is the read-only counterpart. It cannot edit files, and it asks permission before running bash commands. The README positions it for exploring unfamiliar codebases or thinking through a change before committing to it, which is a genuinely useful split. Most agents make you write a careful prompt to get read-only behavior; here it is a keystroke.

There is also a general subagent that handles complex searches and multi-step tasks internally. You can invoke it yourself with @general in a message when you want the agent to fan out on something hairy while you keep working.

This is where OpenCode gets interesting for anyone watching their API bill. It supports 75+ model providers, so you can plug in Claude, GPT, Gemini, or a local model running on your own machine.

If you do not want to configure anything, OpenCode includes free models out of the box. And here is the move that will make the paid tools nervous: if you already pay for ChatGPT or GitHub Copilot, you can just log in with that subscription and use it inside OpenCode. No second subscription, no per-token meter running in the background.

The privacy story is part of the pitch too. The project states plainly that it never stores your code. Your code goes to whichever model provider you chose, which is unavoidable for any AI agent, but OpenCode itself is not keeping a copy. For an MIT-licensed tool you run locally, that is about as good as the trust story gets.

OpenCode started as a terminal tool, and the terminal UI is still its home: a full-screen TUI with the agent, your files, and command history in one place. But it has grown past that.

There is now a desktop app in beta, with downloads for macOS (Apple Silicon and Intel), Windows, and Linux (.deb, .rpm, or AppImage), installable via the releases page or opencode.ai/download. If you live in an editor, OpenCode works there too, so you do not have to context-switch to a separate window to get agent help.

The breadth matters because developers are split on where they want AI. Some want it in the terminal next to their git commands. Some want a standalone app. Some never want to leave VS Code. OpenCode refusing to pick a side is a feature, not indecision.

The fastest path is the install script:

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

Package manager people have options: npm (npm i -g @opencode/cli@latest, also works with bun, pnpm, or yarn), Homebrew on macOS and Linux (brew install anomalyco/tap/opencode is the recommended tap because it stays current), Scoop and Chocolatey on Windows, pacman and AUR on Arch, mise on any OS, and even nix. The README notes you should remove versions older than 0.1.x before installing, which tells you the project moved fast and broke a few things along the way.

The docs live at opencode.ai/docs, and there is an active Discord plus an X account if you want to see what the community is building. The README is translated into 20+ languages, which is a decent proxy for how global the user base has gotten.

Switch if you are paying for a coding agent and the bill bothers you, you want to choose your model per task (or per budget), you like working in the terminal and want the agent where your shell already is, or you care about running open-source tooling you can audit.

Stay where you are if you are deep in Claude Code's ecosystem and happy with it, you rely on a specific Claude Code plugin or workflow that has no OpenCode equivalent, or you would rather pay for one polished integration than assemble your own from 75 providers.

The honest answer is that OpenCode's biggest advantage is not any single feature, it is optionality. The paid agents lock the model, the surface, and the price together. OpenCode unbundles all three: free models to start, your existing subscriptions if you have them, any of 75+ providers if you do not, in a terminal, a desktop app, or your editor. At 216,000 stars and MIT licensed, it is the safest free bet in the AI coding agent space.