Use with OpenCode

OpenCode lets you register any OpenAI-compatible backend as a provider. Otari exposes an OpenAI-compatible endpoint (POST /v1/chat/completions) in both standalone and hybrid modes, so you can route opencode through Otari to get budgets, usage tracking, and traces without changing how you code.

Quick start

Add Otari as a provider in your opencode.jsonc:

{
   "provider": {
     "otari": {
       "npm": "@ai-sdk/openai-compatible",
       "options": {
         "baseURL": "http://localhost:8000/v1",
         "apiKey": "{env:OTARI_API_KEY}"
       }
     }
   }
}

baseURL is the Otari root plus /v1 (opencode appends /chat/completions itself). Point it at the gateway you are actually using: http://localhost:8000/v1 for local standalone development, your self-hosted gateway URL plus /v1 when connected to otari.ai, or https://api.otari.ai/v1 when using otari.ai's hosted gateway.

Export your key so opencode reads it from the environment instead of the config file, then run with a model:

export OTARI_API_KEY=<your-token>  # standalone API key, or tk_ user token opencode --model otari/openai:gpt-4o

The apiKey is sent as Authorization: Bearer <token>, which Otari accepts for both standalone API keys and connected user tokens.

Choosing a model

The provider id is otari (from your config) and the model selector is whatever your deployment expects:

Standalone, any configured provider: provider:model, e.g. otari/openai:gpt-4o, otari/anthropic:claude-sonnet-4-6, otari/mistral:mistral-large-latest.
Connected to otari.ai, managed models: otari/mzai:<catalog-id>, for example otari/mzai:moonshotai/Kimi-K2.6. These run only through otari.ai's hosted gateway.
Connected to otari.ai, your own provider keys: otari/openai:gpt-4o or otari/openai/gpt-4o, plus the equivalent Anthropic or Mistral forms. An mzai: prefix selects the managed catalog, so adding it to a proprietary model misroutes it.

Any model in the catalog works; Otari routes the request to the right provider and records usage and cost for it the same way as any other client.

Image attachments

OpenCode does not read capabilities from the gateway's GET /v1/models. For a custom OpenAI-compatible provider it resolves them from the models.dev catalog, keyed by provider id plus model id, or from an explicit model definition in your config. Because otari is a custom provider id, opencode finds nothing in the catalog, defaults the model to attachment: false, and strips the image before the request leaves your machine. The error surfaces as this model does not support image input.

Declare the capability yourself:

{
   "provider": {
     "otari": {
       "npm": "@ai-sdk/openai-compatible",
       "options": {
         "baseURL": "http://localhost:8000/v1",
         "apiKey": "{env:OTARI_API_KEY}"
       },
       "models": {
         "anthropic:claude-sonnet-4-6": {
           "name": "Claude Sonnet 4.6 (via Otari)",
           "attachment": true,
           "tool_call": true,
           "reasoning": true,
           "modalities": {
             "input": ["text", "image"],
             "output": ["text"]
           }
         }
       }
     }
   }
 }

attachment: true and image in modalities.input are the two keys that matter. With those set, OpenCode sends the image as image_url and Otari passes it through to the provider.

The model key must match what you pass to --model, minus the otari/ provider prefix. You need one entry per model you want to attach images to.

If images still fail after this change, update opencode — some versions have a translation bug where file parts aren't converted to image_url for custom providers (opencode#20802).

Quick start

Choosing a model

Image attachments

See also