# OpenCode review: experiences, strengths and weaknesses | llmnet Apps [← Apps overview](/en/) # OpenCode: review By Ivo Donker · October 28, 2026 — created with AI assistance (writer: DeepSeek V4.1 Flash; critic: gpt-oss-120b via Groq) OpenCode: [Review](/en/opencode-review)[Install & set up](https://apps.llmnet.nl/en/opencode-installatie)[News](/en/opencode-nieuws)[Technical tutorial](https://apps.llmnet.nl/en/opencode-tutorial) Current as of September 2026, compiled from our own notes and news archive; text machine-generated and editorially checked. Check versions and commands in the official documentation. OpenCode is an open-source coding agent that you use in the terminal, as a desktop app or as an editor extension. The big difference from Claude Code, Codex and Cursor: OpenCode is not tied to a single model provider. This review looks at what that delivers, where it grates and who it is for (as of September 2026). ## What it is and who it is for OpenCode is developed by Anomaly and is fully open source. You connect one or more model providers yourself: the big commercial providers, OpenCode's own Zen service, any service with an OpenAI-compatible API, or local models via Ollama and LM Studio. That makes it the natural choice for developers who want to remain model-independent, want traffic to run through their own router or gateway, or want to work (partly) locally and without the cloud. ## How it works In the terminal interface, you use Tab to switch between the Build agent (all tools) and Plan (analysis without changes). You invoke subagents such as Explore and Scout with @ for read-only research. /init creates an AGENTS.md for your project, and you add extra instruction files via instructions. Everything is configured in opencode.json, in layers: global, per project and managed by an organisation. You can refine permissions per tool (allow, ask, deny) by pattern, for example allowing git * and denying rm *. MCP servers, your own agents as markdown files, a headless HTTP server (opencode serve) and a GitHub integration complete it. ## Strengths - Free choice of model. Switching between providers is a line of configuration. Adding your own router or local model is done via the standard package @ai-sdk/openai-compatible. - Open source. You can see what the agent does, can modify it and are not dependent on one company's roadmap or pricing. - Fine-grained permissions. Patterns per command with "last rule wins" are powerful and easy to read. .env files are not read by default. - Your own agents as files. A reviewer that may not change anything, or a documentation agent with a cheap model, is a markdown file with frontmatter that you share with your team via git. - Many ways to run. Terminal, desktop, editor, Docker, an HTTP server and GitHub Actions. ## Weaknesses and risks - Generous default permissions. Most permissions are set to allow by default. If you are used to a strict default, you have to configure it yourself. Do this before first use in an important repository. - The quality depends on your model. OpenCode is only as good as the model you connect. With a small local model, tool calls and longer tasks often go wrong; it delivers its best work with strong commercial models. - You manage more yourself. You arrange keys, providers and model names yourself. That is the price of freedom, and a wrong provider ID produces unclear errors. - opencode serve is dangerous without a password. The server executes commands on your machine; without OPENCODE_SERVER_PASSWORD and without network protection, that is an open door. - Less built-in sandboxing. Where Claude Code and Codex include an OS-level sandbox, OpenCode relies mainly on permission rules. For truly unattended work, a container or VM is sensible. ## Price and licence OpenCode itself is free and open source. You only pay for the models you use: via your own API keys with the providers, via Zen, or nothing if you run locally. That often makes it the cheapest route, especially in combination with a router that sends routine work to cheap models. ## Alternatives - Claude Code: stronger built-in security (sandbox, hooks), but only Claude models and not open source. - Codex: open source with a good sandbox, but focused on OpenAI models. - Cursor: for those who prefer working in a graphical editor. - Hermes Agent: open source, focused on long-running, self-hosted agents with their own memory. ## Conclusion OpenCode is the best choice for anyone who wants to stay in control of models and infrastructure: multiple providers, your own router, local models or a fully self-hosted setup. You get a lot of flexibility and an open codebase, in exchange for more configuration work of your own and a broader default setting that you have to tighten deliberately. Set the permissions tight first, choose a strong model for the real work, and OpenCode can keep up well with the commercial agents in day-to-day practice.