anomalyco/opencode
118,000 stars in under a year. That's not a typo. OpenCode went from zero to the most-starred AI coding agent on GitHub faster than any developer tool in memory -- and it did it by being the one thing Cursor and Claude Code refuse to be: completely open source, completely provider-agnostic, and completely free. Here's the pitch that mass-converted developers: bring your own model, bring your own keys, keep your own data. OpenCode doesn't care if you're running Claude Opus, GPT-4.1, Gemini 2.5, or a local Llama instance through Ollama. It treats every provider as a first-class citizen. If you've ever felt locked into Anthropic's pricing because Claude Code only works with Claude, or locked into Cursor's $20/month because switching means losing your workflow -- OpenCode is the exit door 5 million developers already walked through. The architecture is what makes it stick. OpenCode runs a client/server split -- the AI agent runs as a background server while the TUI, desktop app, or IDE extension connects as a client. That means you can run the agent on a beefy remote machine and code from a thin laptop over SSH. Try doing that with Cursor. Two built-in agents handle different workflows: a build agent for writing and modifying code, and a read-only plan agent for exploring codebases without accidentally changing anything. There's also a general subagent that handles complex multi-step searches. LSP integration gives it real code intelligence -- not just pattern matching, but actual type-aware navigation and diagnostics. The release velocity tells its own story: 731 releases, 10,045 commits, v1.2.21 shipped March 7, 2026. The team at Anomaly (the same folks behind terminal.shop) ships daily. MCP support means you can extend it with the same server ecosystem Claude Code uses. Install takes one curl command. You're coding in 30 seconds.
Why It Matters
The AI coding tool market just split into two camps: proprietary tools that lock you into one model provider and one pricing tier, and OpenCode, which lets you swap between any model mid-session without losing context. For teams running multiple AI providers -- Claude for complex reasoning, GPT-4.1 for speed, local models for sensitive code -- OpenCode is the only terminal agent that treats this as a first-class workflow rather than a hack. The 118K star count isn't vanity. It represents mass developer frustration with vendor lock-in crystallized into a single tool. When Cursor raised prices and Claude Code required an Anthropic subscription, OpenCode absorbed the exodus. The MIT license means your company can fork it, modify it, and deploy it internally without legal review. If you're building AI-assisted development workflows and don't want your toolchain held hostage by any single company's pricing decisions, OpenCode is the insurance policy that also happens to be the best terminal coding agent available.