The coding agent where you pick the brain
opencode is an open source AI coding agent that runs in your terminal. It writes the code, runs it, and fixes what breaks. The part that matters: it is not tied to one AI company, so you choose which model does which job.
curl -fsSL https://opencode.ai/install | bashOn Windows use scoop install opencode, choco install opencode or npm install -g opencode-ai. The docs recommend WSL for the best experience.
The numbers, straight off the repo
Read on 8 September 2026. The project’s own homepage still says 195K stars, so these came off GitHub itself, which is ahead of it.
205.8k stars · MIT license · 1,003 contributors · 871 releases, the latest four days before this page was written. That last one is the tell: this is a repo people are still shipping to every week.
You describe it. It builds it.
You type what you want in plain English. A tool, a script, a whole app. It reads your files, writes the code, runs it, and when something breaks it reads the error and fixes it.
It ships as a terminal app, a desktop app, and an IDE extension. Two built-in agents: build has full access, plan is read-only for when you just want it to think.
Route the job to the right brain
This is the reason to care, and it is not the price. Most coding agents hand you one model for everything: the architecture question and the rename-this-variable question go to the same expensive brain.
opencode connects to 75+ providers through Models.dev, and that list includes models that run entirely on your own machine through Ollama or LM Studio. No API key, no request leaving your laptop.
So the split looks like this
Hard problems to a frontier model. Architecture, tricky debugging, anything where being wrong is expensive.
Everything else to a free local model. Renames, boilerplate, formatting, the long tail of small edits.
Switch inside the app with /models. Set a default in opencode.json.
Every way to install it
All of these are from the official docs. Pick one.
Anywhere
curl -fsSL https://opencode.ai/install | bashNode
npm install -g opencode-aimacOS and Linux
brew install anomalyco/tap/opencodeWindows
scoop install opencodeAlso choco install opencode. Bun support on Windows is still in progress, and the docs suggest WSL if you want the smoothest run.
Arch
sudo pacman -S opencodeYou will need an API key for any cloud provider you want to use. For a fully local setup, install Ollama first and point opencode at it.
Not one person’s side project
The thing that kills most open source tools is that they stop. This one has 1,003 contributors and a commit graph that has not gone quiet.
MIT licensed, so you can read it, fork it, and use it commercially. The repo moved from sst/opencode to anomalyco/opencode, and the old link still redirects.
The honest bits
Free does not mean zero cost for everything. The tool is free and MIT licensed, and there are free models included. But if you point it at Claude or GPT, you pay that provider’s rates. The saving comes from routing, not from magic.
Local models are not frontier models. A 30B model on your laptop is genuinely useful for edits and boilerplate. It is not going to reason like the best cloud model, and pretending otherwise wastes your time.
It is not a Claude Code clone. I have seen it described that way and the project never claims it. It is its own agent with its own design. Judge it on that.
Windows is the rough edge. It works, but the docs themselves point at WSL. If you are on Windows and something feels off, that is where to start.
Start here
Install it, run /models, and set up one cloud model plus one local one. That single change is the whole idea of this reel.
Every number on this page was read off github.com/anomalyco/opencode and opencode.ai on 8 September 2026. They move. Check the repo for today’s.