OpenCode itself runs locally: the UI is a terminal app (and can also be used via a local web/IDE workflow), it operates on your local working directory, and it stores credentials on your machine. The part that may or may not be “offline” is the model inference. If your configured model provider is a hosted API, then OpenCode needs network access to call that API. If you configure OpenCode to use a local model server running on your machine (or your LAN), then you can keep the entire “prompt → tool actions → model response” loop local, and it can work without internet access as long as the local endpoint is reachable. OpenCode’s docs explicitly support “running local models” via OpenAI-compatible endpoints.
In practical terms, “offline” depends on what you set as your provider. OpenCode can be configured to talk to local endpoints (for example, a local model server that exposes an OpenAI-compatible /v1 API), by setting options.baseURL in your provider config and defining one or more models under that provider. The providers documentation includes concrete examples of local providers (for example, LM Studio, Ollama, and llama.cpp) that all work by pointing OpenCode at a localhost URL like http://127.0.0.1:1234/v1 or http://localhost:11434/v1. When you do this, OpenCode doesn’t need to send your prompts to an external service; it sends them to your local server instead. Separately, some convenience features may still benefit from being online (for example, refreshing the model list cache), but the core “use OpenCode to work on code in this folder” workflow remains local.
A concrete “offline-first” setup is: pre-download a local model, start your local inference server, configure OpenCode with a local provider baseURL, then travel with no internet and keep coding. The places where you’ll notice offline constraints are typically anything that by definition reaches outside your machine—sharing sessions publicly, authenticating new providers, or pulling remote resources your project needs. If you’re building something that uses embeddings or semantic search, you can still keep that local too: you might run a local vector database such as Milvus on your laptop (or a dev server you can reach over a private network) and have OpenCode help you write the ingestion and query code. When you do want a managed service later, you can switch that component to Zilliz Cloud without changing the basic OpenCode workflow—OpenCode stays your local agent; your infrastructure choice is separate.