OpenCode Just Passed 160,000 GitHub Stars. The Most Popular Coding Agent in the World Is MIT-Licensed and Brings Its Own Keys. Here's When That Actually Changes Your Decision.
OpenCode crossed 160,000 GitHub stars, which makes it the most-starred open-source AI coding agent by a wide margin. It's built by the team behind SST (now operating as Anomaly), it's MIT-licensed, it runs in your terminal, and it connects to 75-plus model providers with your own API keys. Those are the facts. The question worth your time isn't whether it's popular. Popularity is a vanity metric for tools the same way it is for tweets. The question is whether "open and model-agnostic" is worth the switching cost for the way you actually work.
I'm writing this the same week Google is shutting off Gemini CLI for consumer accounts and pushing everyone to a closed-source replacement. That timing is useful, because it turns an abstract benefit ("model-agnostic") into a concrete one. The value of bringing your own keys is that nobody gets to deprecate your workflow on a date you didn't pick.
What OpenCode actually is
OpenCode is terminal-native, not an IDE and not a web app. It reads your code through Language Server Protocol integration, so the model gets real type information, function signatures, and compiler diagnostics instead of just raw text, across TypeScript, Python, Rust, Go, and a long tail of other languages. It supports multiple sessions, so you can run more than one agent against the same project at once. And the defining design choice: it's model-agnostic. You point it at whatever provider you want (Anthropic, OpenAI, Google, a local model) and pay that provider directly for tokens.
The backstory matters for trusting it. This isn't a weekend project that got lucky. The SST/Anomaly team has been shipping serverless infrastructure tooling for years and has real backing. OpenCode launched in 2025 and reached this scale in under a year. That's fast, but it's fast on top of a team that knows how to maintain software, which is the part that determines whether an open-source tool is still alive in two years.
Why model-agnostic is the actual pitch
Claude Code is excellent and tied to Anthropic. Cursor is excellent and is a subscription IDE. Both are good products, and for a lot of people the right answer is to just use one of them and stop thinking about it. I'm not here to talk you out of tools that work.
But there's a category of risk that only shows up when a vendor changes the terms, and this week is a live demonstration of it. A tool locked to one provider inherits that provider's decisions: pricing changes, model deprecations, rate limits, a CLI that stops working on June 18. A model-agnostic tool turns those from emergencies into config edits. If your provider doubles prices or sunsets the model you depend on, you change a line and keep working. That's not a feature you use every day. It's insurance you're glad you had the one day you need it.
The honest framing is that this is a hedge, not a religion. OpenCode being open-source doesn't make it better at writing code than Claude Code. It makes it harder to strand you. Those are different properties, and you should buy the second one deliberately, not because a repo has a lot of stars.
When switching is worth it, and when it isn't
Switch if you're already feeling vendor risk: you've been burned by a deprecation, you run real automation that can't afford a surprise, or you want to route different tasks to different models without rebuilding your setup each time. Switch if you value reading and modifying the tool itself, because MIT-licensed and inspectable is a real difference from a closed binary.
Don't switch on reflex. If Claude Code or Cursor fits your brain and your workflow, the polish and the sane defaults are worth money, and "worth money" is the right way to think about it, because the cost of moving is real. You'll spend a few days relearning muscle memory, reconfiguring keys, and rediscovering which defaults you'd taken for granted. A solo operator's scarcest resource is uninterrupted building time, and a tool migration spends exactly that. Don't spend it to chase a star count.
What I'd actually do
Install OpenCode this weekend and run it against a real project (not a toy, something with your actual mess in it) pointed at the same model you already pay for. You're not committing. You're pricing the switch: how much does your daily loop change, what breaks, what's better. Two hours tells you most of it.
Then make the call on one question, not on hype: how exposed am I if my current coding tool's vendor changes the deal? If the answer is "barely, I use it interactively and I'd shrug," stay where you are and ignore the stars. If the answer is "a lot, it's in my automation, or I'm one pricing email from a bad month," then model-agnostic stops being a buzzword and becomes the reason to keep a working OpenCode setup in your back pocket, configured and ready, whether or not it's your daily driver. The week a megacap deprecates a free CLI on 30 days' notice is a cheap week to learn that lesson on someone else's tool instead of your own.
Author
Lukas
@lukcombinator