A year ago, Jay V stood in front of 30 developers at a DevTools Toronto meetup and showed them a terminal-based coding agent. No launch party. No Product Hunt campaign. Just a GitHub repo and a curl command. Today, OpenCode has 8 million monthly active users, and the company behind it disclosed a projection of $25 million in expected annual revenue in a June 2026 interview. That is a faster growth curve than almost any developer tool in recent memory, and it happened without the thing most AI tools rely on: a captive model.
What OpenCode actually is
OpenCode has become one of the most-starred open-source developer tools of 2026 by doing something deliberately unglamorous: giving developers a free, model-agnostic agent that runs where they already work and doesn't lock them into a vendor.
It is a terminal-based AI coding agent built by the team behind SST (now Anomaly). It runs locally on your machine and connects to 75+ AI providers, giving you full control over which models process your code.
The architecture is the point. The core design decision is separation of concerns: the agent orchestration layer and the model layer are fully decoupled. That is not a marketing line - it has a concrete consequence. The company maintains models.dev, described as the largest catalog of AI model metadata, with community-contributed integrations for new providers. When a lab releases a new model, contributors usually have an OpenCode integration open within hours.
The project ships two built-in agents switchable with a Tab key. Build is the default, full-access agent for development work. Plan is a read-only agent for analysis and code exploration. There is also a general subagent for complex searches and multi-step tasks. Sessions are stored locally in SQLite, conversations never leave the machine unless you are hitting a cloud API yourself.
The two technical decisions that separate it from everything else
The first is LSP integration. Most coding agents work by pasting your code into a context window. OpenCode integrates with the Language Server Protocol to feed compiler diagnostics, type errors, and import resolution data directly to the model as structured tool results.
OpenCode spawns Language Server Protocol servers and feeds compiler diagnostics back to the model after every edit. When the model introduces a TypeScript type error, it sees the error and self-corrects. This is unique to OpenCode in 2026 - commercial tools like Claude Code do not have LSP integration for real-time error feedback.
The second is Git-native state management. Every meaningful change creates a snapshot automatically. You roll back with /undo; no manual commits required. This matters because the alternative - trusting the model to reverse its own broken changes - fails more often than the demos suggest.
For environments with strict data governance requirements, OpenCode supports fully local execution. Using Ollama or LM Studio to host an open-weight model, users can run OpenCode without any network calls to cloud services. Ollama exposes an OpenAI-compatible API that OpenCode consumes as a standard provider. This capability is marketed as Air-gapped Mode and targets developers in regulated industries such as defense, healthcare, and financial services.
The Anthropic dispute and what it reveals
OpenCode, the most-starred open source coding agent, blew past 165k GitHub stars and then got into a public fight with Anthropic over subscription login.
After a January 2026 dispute, Anthropic forced OpenCode to remove Claude Pro/Max subscription login. You can still use Claude with a raw API key, and OpenAI now partners with the project so ChatGPT subscriptions work natively.
The episode is instructive. Anthropic's concern, presumably, was that OpenCode's subscription passthrough let users get Claude access in a context Anthropic didn't fully control. But the fallout only underscored why the model-agnostic design matters: OpenCode's users shrugged, pointed at a different provider, and the star count kept climbing. A tool that depends on one model's goodwill is brittle. One that treats models as interchangeable infrastructure is not.
Where it actually falls short
Stars are not a reliability audit. A few honest gaps worth knowing:
Browser automation is absent. Codex CLI ships a Chrome extension for browser-driven tasks. OpenCode doesn't have an equivalent, which matters for full-stack web development where frontend interaction is part of the loop.
Where it struggles: OpenCode involves more setup steps. Provider configuration, API key management, model selection - overhead that Claude Code and Codex eliminate by design. The desktop app is still in beta. Documentation lags the feature set.
The business model carries its own tension. The free MIT core is a durability question. A large enough developer community can fork an MIT project and route around paid tiers completely. Anomaly's long-term monetization depends on whether subscribers stick around once they understand the underlying architecture. At $10/month for Go, the price is low enough that most won't bother self-hosting - but the option exists, and that overhang is real.
What the ecosystem around it looks like
OpenCode didn't emerge into a stable market. Google announced it's retiring the 104k-star Gemini CLI, replacing it with a closed-source successor. Block handed Goose to the Linux Foundation. Roo Code, at 24k stars, archived itself.
And a brand-new harness called Pi - from Armin Ronacher, the person behind Flask and Jinja2 - showed up and crossed 54k stars by running on a system prompt under 1,000 tokens.
That flux is the context for OpenCode's dominance. When commercial tools shift billing models overnight - GitHub Copilot moved to usage-based AI credits on June 1, at $0.01 per credit, retired premium requests, and paused new paid sign-ups during the rollout
- developers look for something they can reason about and control. An MIT-licensed tool where the harness and the model are separate layers is a different kind of bet than a SaaS product that can reprice next quarter.
172,000 stars built up in one year, on an MIT-licensed terminal tool with no vendor lock-in, is a different kind of signal than a proprietary tool with built-in distribution. OpenCode got here by making the agent layer cheap to adopt and the provider choice completely open.
The honest answer to whether that holds: it depends on whether Anomaly keeps shipping. Aider's last repo push was May 22, 2026, slower than OpenCode and Cline which push daily. Commit cadence is the only metric that predicts whether a coding agent is a tool or a graveyard. For now, OpenCode is shipping daily, and a Beagle-style agent that needs to invoke a coding harness as part of a Slack workflow has a clear first choice.