docker/sbx-kit-opencode-model-runner repository overview

⁠opencode-model-runner

A fork of the built-in opencode agent that routes all model API calls to a local Docker Model Runner⁠ instance via its OpenAI-compatible endpoint. Useful for offline development, cost-free experimentation, or testing custom local models with the OpenCode UI.

A mixin variant lives in ../opencode-model-runner-mixin⁠, for layering the same agent and configuration onto a shell base instead.

⁠Usage

sbx run "docker.io/docker/sbx-kit-opencode-model-runner:latest" ~/my-project

Or from a git URL or a local clone of this repo:

sbx run "git+https://github.com/docker/sbx-kits-contrib.git#dir=opencode-model-runner" ~/my-project
sbx run ./opencode-model-runner/ ~/my-project

The workload reference passed to sbx run identifies the kit. Its descriptor provides the matchable name opencode-model-runner (opencode-model-runner.yaml⁠).

All models available via docker model ls are automatically discovered and selectable in OpenCode via /models.

⁠How it works

OpenCode reads its provider configuration from ~/.config/opencode/opencode.json. This kit uses a lifecycle@1 capability's files: entry to drop that JSON into the sandbox at startup, declaring:

• An @ai-sdk/openai-compatible provider (dmr) whose baseURL is http://host.docker.internal:12434/v1 (Model Runner's OpenAI-compatible endpoint), with modelsDiscovery.enabled: true so the provider queries Model Runner's /models endpoint at startup.
• The opencode-models-discovery plugin (with smartModelName: false) which surfaces the discovered models in OpenCode's model picker.

Any model pulled with docker model pull appears automatically — no manual config edits needed.

⁠Related

• Docker Model Runner⁠
• OpenCode with Docker Model Runner for Private AI Coding⁠, the inspiration for this kit