Try OpenCode
Overview
OpenCode is the coding agent that grew fastest in 2026 by refusing the industry's default bargain. Every commercial agent ties you to its vendor's models; OpenCode ships the agent and lets you bring your own key — Claude, GPT, Gemini, or a model running on your own hardware. For teams with existing model contracts, procurement constraints, or simply a preference for not being resold inference at a markup, that is the entire argument.
It reached over 120,000 GitHub stars, which puts it among the most-starred developer tools of the year, and the growth is not purely ideological. It runs in the terminal, which means it works over SSH, inside containers, on a build server and in every editor at once, rather than being a feature of one IDE. Developers who live in tmux rather than in a GUI found something built for them.
The commercial layer arrived without breaking the free core: Zen usage-based plans for people who would rather buy credits than manage provider keys. The project is maintained by SST, and the governance question that follows every fast-growing open-source developer tool — what happens when the commercial layer and the free core disagree — is worth watching rather than worrying about yet.
Key Features
Terminal-Native
Runs where developers already work. Works over SSH, in containers and on remote build machines, with no dependency on a particular editor.
Bring Your Own Model
Point it at Claude, GPT, Gemini or a locally hosted model. You control the cost, the provider and where your code is sent.
Fully Open Source
The agent loop is auditable, which matters for organisations that must know what leaves their network and why.
Multi-File Codebase Work
Plans and executes changes across many files, rather than completing the line you are typing.
Zen Usage Plans
Optional paid credits for teams that would rather not manage provider keys and rate limits themselves.
Self-Hostable End to End
Combined with a local model, code never leaves your infrastructure — a genuinely rare property among capable coding agents.
Pros & Cons
Advantages
- No vendor lock-in: your model, your key, your data path
- 120,000+ stars means a very active ecosystem and fast bug fixes
- Terminal-native, so it works over SSH and in CI where GUI tools cannot
- Free core, with optional paid credits rather than a paywall
- Can run fully offline against a local model
Disadvantages
- You manage model keys, rate limits and cost yourself
- No integrated GUI — a real barrier for editor-centric developers
- Quality depends entirely on which model you point it at
- Governance of the free core versus the commercial layer is unproven
Pricing Plans
| Plan | Price | Key Features |
|---|---|---|
| Open Source | Free | Full agent, bring your own model API key |
| Zen | Usage-based | Prepaid credits, no provider key management |
Best Use Cases
OpenCode Excels At:
- Teams with existing model contracts or procurement constraints
- Remote development over SSH and inside containers
- Regulated environments needing a fully self-hosted coding agent
- Developers who want to read the agent loop they are running
May Not Be Ideal For:
- Developers who want a polished GUI experience
- Teams that want one vendor to call when something breaks
- Anyone unwilling to manage model costs
How It Compares
OpenCode vs Claude Code
Claude Code is the stronger out-of-the-box agent, tuned end to end against Anthropic's models and backed by a vendor. OpenCode trades that polish for freedom: any model, any environment, full source. Teams that must control the data path or already buy inference elsewhere pick OpenCode.
OpenCode vs Cursor
Cursor is an editor with an agent inside it; OpenCode is an agent with no opinion about your editor. If you work in a GUI, Cursor is more comfortable. If you work over SSH or across many machines, Cursor cannot follow you there.
Final Verdict
Our Recommendation
OpenCode earned its 120,000 stars on a straightforward proposition: the coding agent should not decide which model you use or where your code goes. Terminal-native execution makes it the only serious option for a lot of real environments — remote servers, containers, air-gapped networks — and the ability to run it against a local model is a genuine differentiator for regulated work. It is less polished than the commercial agents and you carry the cost management yourself. If you value control over convenience, this is the coding agent to adopt; if you want it to just work on day one, buy one instead.