OpenCode can connect to a wide range of models because it’s built around a provider system that supports many LLM backends (including cloud APIs and local runtimes). In the OpenCode docs, model support is described in terms of “providers” and “models”: you connect a provider via /connect, OpenCode stores the credentials locally, and then you select a model using /models or by setting a default model string in config. The key technical point is that OpenCode is designed to be provider-agnostic and uses a standard integration layer so it can support a large catalog of providers, not just a single vendor. That also means you can switch models per project, per session, or even per command, which is useful when you want one model for fast iteration and another for deeper reasoning or longer context tasks.
In practical terms, OpenCode commonly connects to major hosted providers (for example, OpenAI, Anthropic, Google, and cloud marketplaces such as Amazon Bedrock and Azure OpenAI) and to routing layers (for example, OpenRouter). It also supports running local models by treating them as “OpenAI-compatible” endpoints via a configurable baseURL, which fits tools like Ollama or LM Studio when they expose an OpenAI-style API. Configuration is explicit and developer-friendly: you can define providers in opencode.jsonc, specify model IDs, set per-model options, and define variants (for example, “high reasoning” or “max output”) that you can cycle at runtime. This matters when you care about reproducibility: if your repo has a known-good model setting and temperature/options, you can store that in the project config so teammates get similar behavior without manual tuning. The /connect flow keeps secrets out of repos by storing credentials in a local auth file rather than asking you to commit keys into config.
OpenCode’s model connectivity also makes it easy to integrate specialized gateways. For example, the Z.AI developer documentation shows how to configure OpenCode to use Z.AI’s GLM models, which is another instance of the same provider pattern: authenticate, select the provider, and set the desired model in config or via the UI. Once you have this flexibility, you can build higher-level workflows on top of it, especially retrieval-augmented coding: index your internal docs, code snippets, or past incident notes, retrieve the right context for the current task, and attach it to the prompt before the model is called. If you implement that retrieval layer with a vector database such as Milvus or Zilliz Cloud, you can keep OpenCode sessions lightweight while still giving the agent “memory” that scales to large organizations and long timelines. The model remains replaceable, the retrieval remains reusable, and OpenCode becomes the glue that orchestrates “select model → fetch context → generate patch → review/apply changes” in one consistent developer workflow.