OpenCode Crossed 120K GitHub Stars — and Even Anthropic’s Legal Threats Couldn’t Slow It Down

Eight months. That’s how long it took OpenCode to go from a quiet launch by a serverless framework team to becoming the most-starred open-source AI coding agent on GitHub. With over 120,000 stars, 800+ contributors, 10,000+ commits, and more than 5 million monthly active developers, the numbers alone tell a story. But the real story is messier — and more interesting — than a simple growth chart.

OpenCode landed on Hacker News again on March 20, 2026, pulling 376 points and 180+ comments. InfoQ, The New Stack, and others have run features on it. And along the way, it got into a very public fight with Anthropic that somehow made it even more popular.

From Serverless Framework to AI Coding Powerhouse

OpenCode didn’t come out of a well-funded AI lab. It came out of Anomaly, a small developer tools company best known for SST (Serverless Stack), an open-source framework for building apps on AWS. Jay V (CEO), Frank Wang (CTO), and Dax Raad — an early SST user who became a co-founder — had been building infrastructure tools together since 2021, when they went through Y Combinator and raised backing from the founders of PayPal, LinkedIn, Yelp, and YouTube.

The pivot to AI coding happened when Jay noticed something obvious in hindsight: as the number of capable coding models exploded, developers would resist being locked into a single provider’s tool. Claude Code requires Anthropic models. GitHub Copilot is tightly coupled to OpenAI. Jay bet that an open-source, model-agnostic agent would find a massive audience.

On June 19, 2025, OpenCode launched. Within five months, it had 650,000 monthly active users. By March 2026, that number crossed 5 million.

No Lock-In: The Core Philosophy

The single biggest differentiator is model freedom. OpenCode supports 75+ models through Models.dev — Claude, GPT, Gemini, Groq, AWS Bedrock, Azure OpenAI, OpenRouter, and local models via Ollama or LM Studio. Developers bring their own API keys (the BYOK model) and pay only provider rates. No markup, no subscription required for the core tool.

Beyond model choice, OpenCode ships in three form factors:

Terminal — The original and still primary interface. Built with Bubble Tea, it offers an interactive TUI with vim-like editing, streamed markdown output, and multi-session management backed by SQLite. You can switch between a “build” agent (full access, writes and executes code) and a “plan” agent (read-only, for analysis) with the Tab key.
Desktop app — A standalone application for developers who prefer a windowed experience.
IDE extension — Available for VS Code, Cursor, and any editor supporting ACP (Agent Client Protocol), including JetBrains, Zed, Neovim, and Emacs. Hit Cmd+Esc on Mac (Ctrl+Esc on Windows/Linux) to open OpenCode in a split terminal view with automatic context awareness.

Under the hood, OpenCode integrates Language Server Protocol (LSP) servers for Rust, Swift, Terraform, TypeScript, PyRight, and more — giving the AI access to type information, diagnostics, and code structure that raw file reading misses. It also implements the Model Context Protocol (MCP) for extending capabilities through external tools and services.

Privacy is another selling point. Anomaly says OpenCode does not store code or context server-side. Session sharing is opt-in, with granular controls: manual sharing, auto-sharing, or fully disabled. For teams working on sensitive projects, sharing can be locked off at the configuration level.

The Anthropic Showdown

Here’s where the story gets spicy.

In late 2025 and early 2026, OpenCode (and similar tools) figured out they could route requests through Anthropic’s internal Claude Code APIs instead of the standard Claude API. The practical effect: developers with a $20/month Claude Pro subscription could use Claude’s best models through OpenCode at subscription rates, bypassing the much more expensive per-token API pricing.

Anthropic did not appreciate this.

On January 9, 2026, Anthropic deployed server-side blocks that prevented third-party tools from authenticating with Claude Pro and Max subscription OAuth tokens. On February 19, they updated their terms of service to explicitly ban this practice. And they sent legal notices to OpenCode, forcing the team to remove Claude Pro/Max account key support from the codebase.

The community reaction was fierce. Ruby on Rails creator DHH publicly criticized Anthropic, pointing out the irony of “a company built on training models on our code” blocking open-source tools from accessing those models. Hacker News threads debated whether Anthropic’s walled-garden approach would backfire.

OpenAI saw an opportunity and seized it. They officially partnered with OpenCode and extended subscription support — a deliberate signal that while Anthropic was closing doors, OpenAI was opening them. The drama ultimately drove more attention and users to OpenCode, not fewer.

How OpenCode Stacks Up Against the Competition

The AI coding agent market in 2026 is crowded. Here’s where OpenCode sits relative to the key players:

Claude Code scores 80.9% on SWE-bench and offers deep reasoning with Opus 4.6’s 1M token context window. It recently shipped Agent Teams for multi-agent coordination, MCP server integration, and custom hooks. The trade-off: it only works with Anthropic models. If you’ve been following the Claude Code ecosystem, you know it’s powerful — but the lock-in is real.

GitHub Copilot is the $10/month safety net that works everywhere — VS Code, JetBrains, Visual Studio, Xcode, Neovim. But multi-file editing remains less reliable, and its agent mode still trails behind dedicated agentic tools.

Cursor has gone from IDE wrapper to building its own models. Composer 2 scores 61.7% on Terminal-Bench 2.0 at a fraction of Opus’s cost. But Cursor is IDE-first, not terminal-first.

Cline is OpenCode’s closest open-source competitor. Also model-agnostic, also available in the terminal, also supports ACP. The key difference is ecosystem momentum: OpenCode’s 120K stars and plugin ecosystem (tools like oh-my-opencode that spin up multi-agent squads from a single instance) give it a larger community and more third-party tooling.

The honest recommendation from most developers who’ve tried multiple tools: use more than one. OpenCode or Claude Code in the terminal for hard problems, Cursor or Windsurf as your daily IDE agent, and Copilot as the universal fallback.

OpenCode Go: The $10/Month Subscription Play

While the core tool is free and BYOK, Anomaly launched OpenCode Go for developers who want a simpler on-ramp. Priced at $5 for the first month and $10/month after, it provides access to capable open-source models — GLM-5, Kimi K2.5, MiniMax M2.5, and MiniMax M2.7 — without needing to set up API keys.

Usage limits are dollar-equivalent rather than request-based: $12 every 5 hours, $30 weekly, $60 monthly. Because MiniMax M2.5 has extremely low per-request costs, that monthly cap translates to roughly 100,000 requests for that model. Other models like GLM-5 have lower effective request limits.

Models are hosted in the US, EU, and Singapore for global coverage. OpenCode Go is currently in beta and clearly aimed at the international developer market — particularly those in regions where setting up API billing with US-based providers is friction-heavy.

It’s not going to replace a Claude or GPT subscription for developers who need frontier-level reasoning. But for everyday coding tasks, quick prototyping, or developers working on tighter budgets, it fills a real gap.

FAQ

Is OpenCode free to use?
Yes. The core OpenCode tool — terminal, desktop, and IDE extension — is fully open-source and free. You bring your own API keys for whichever model provider you prefer and pay only their standard rates. There’s no markup. OpenCode Go ($10/month after a $5 intro month) is an optional subscription that provides access to open-source models without needing to manage API keys.

What models does OpenCode support?
Over 75 models through Models.dev. This includes Claude (Anthropic), GPT (OpenAI), Gemini (Google), Groq, AWS Bedrock, Azure OpenAI, OpenRouter, and local models via Ollama or LM Studio. OpenCode recommends certain models but does not force any specific provider — model choice is entirely up to the developer.

How does OpenCode compare to Claude Code?
Claude Code offers deeper reasoning (80.9% SWE-bench) and features like Agent Teams and a 1M token context window, but only works with Anthropic’s models. OpenCode supports 75+ models from any provider, is fully open-source, and gives developers complete control over their data. The trade-off is that Claude Code may produce higher-quality outputs on complex tasks when using Opus 4.6, while OpenCode provides flexibility and avoids vendor lock-in.

Is OpenCode safe for proprietary/private codebases?
OpenCode’s privacy architecture means code and context are not stored on Anomaly’s servers. All processing happens through the model provider you choose. Session sharing is off by default and can be fully disabled at the team level. For maximum privacy, developers can run local models through Ollama, keeping everything on their own hardware.

Can I use OpenCode in my IDE instead of the terminal?
Yes. OpenCode has an official VS Code extension and supports the Agent Client Protocol (ACP), which means it works with JetBrains IDEs, Zed, Neovim, Emacs, and other compatible editors. There’s also a standalone desktop app for those who prefer a windowed interface outside the terminal.

Top AI Product

Leave a comment Cancel reply