OpenCode Zen¶

Zen supports completions through OpenCode's Chat Completions and Messages routes. It owns conversation history and selects the payload format from the active model ID. Embeddings and decisions aren't supported.

Connection¶

import intuit;

auto router = new Zen("your-zen-key");
ModelDetails[string] catalog = router.catalog();

The constructor accepts key, url (default https://opencode.ai), and display name (default "OpenCode Zen"). A nonempty key uses bearer authentication. Messages requests also set anthropic-version to 2023-06-01.

The base URL is used as supplied. Requests append /zen/v1/models, /zen/v1/chat/completions, or /zen/v1/messages.

Model Selection¶

The current implementation uses model ID prefixes, not catalog API-family metadata.

Model ID	Configuration and route
Starts with `claude-`	`ClaudeModelConfig`, Messages API.
Starts with `qwen`, except `qwen3-coder`	`ClaudeModelConfig`, Messages API.
Starts with `gpt-`, `gemini-`, or `jev-`	Rejected when requesting a completion.
Other IDs, including `qwen3-coder`	Base `ModelConfig`, Chat Completions API.

Activation loads the catalog once if needed and accepts IDs absent from it. Unsupported prefixes can therefore be activated but still fail when a completion is requested. A model appearing in the catalog doesn't guarantee that its API family is supported by this adapter.

For a supported model, select its exact ID before making requests.

router.active("your-supported-model-id");
Completion result = completions(router, "Explain associative arrays.");

Activation preserves history and updates the compactor limit only when the catalog reports a positive context length. Changing API families doesn't translate existing history into a different message format.

Configuration and Limitations¶

Use Models for base settings and Anthropic for ClaudeModelConfig settings and message-translation limits. The Messages path has the same shared tool-history limitation described there.

The Chat Completions path uses the base parser, not Qwen XML extraction. Zen doesn't implement the Responses API, Gemini-native requests, decisions, embeddings, or streaming.

Catalog Mapping¶

The catalog currently supplies only model IDs, names, descriptions, and positive context lengths. Prices, capabilities, modalities, and output limits aren't mapped, even if the service returns them.

See Model Catalogs for shared metadata and Getting Started for router conversation behavior.