OpenCode Local Ollama Provider

Native Ollama model discovery for OpenCode.

opencode-local-ollama is a native OpenCode provider plugin. It discovers the models you already have in Ollama and registers them as ollama/<model> in OpenCode, so they show up in /models without a manual sync step.

Why

Existing Ollama/OpenCode helpers usually generate config files or require a separate sync command. This plugin runs during OpenCode startup and uses OpenCode's provider hook directly.

Features

• Discovers local models from Ollama on OpenCode boot.
• Uses /api/tags to list local models.
• Uses /api/show for exact model metadata.
• Registers models under the ollama provider.
• Uses Ollama's OpenAI-compatible endpoint at /v1.
• Respects OLLAMA_HOST and plugin options.
• Model agnostic: no hardcoded Qwen, Llama, Gemma, Mistral, or DeepSeek rules.
• Strict by default: no guessed capabilities.
• Skips non-chat models when Ollama does not report completion or chat capability.

Install

npm install -g opencode-local-ollama

Then add it to your OpenCode config:

{
   "plugin": ["opencode-local-ollama"]
}

Restart OpenCode and run:

/models

You should see entries like:

ollama/llama3.2:latest
ollama/gemma4:31b
ollama/qwen3.5:35b

Usage

Pull a model with Ollama:

ollama pull llama3.2

Restart OpenCode. The model becomes available as:

ollama/llama3.2:latest

Configuration

Default behavior uses OLLAMA_HOST if set, otherwise http://localhost:11434.

{
   "plugin": [
      [
        "opencode-local-ollama",
        {
          "host": "http://localhost:11434"
        }
      ]
    ]
}

Options:

| Option | Default | Description | | --- | --- | --- | | host | OLLAMA_HOST or http://localhost:11434 | Ollama API host. May include a proxy path. | | baseURL | unset | Alias for host. If it ends in /v1, the plugin derives the Ollama API base from it. | | providerID | ollama | OpenCode provider ID to register. | | timeout | 5000 | Ollama request timeout in milliseconds. | | context | 4096 | Explicit context limit to expose to OpenCode. | | output | 4096 | Explicit output token limit to expose to OpenCode. | | useModelInfoContext | false | Use GGUF architecture context from model_info when num_ctx is not configured. Disabled by default because it can be larger than the runtime context Ollama actually uses. |

Capability Mapping

The plugin treats Ollama as the source of truth.

| Ollama capability | OpenCode mapping | | --- | --- | | completion or chat | text input/output model | | tools | tool calling enabled | | vision or image | image attachments enabled | | thinking or reasoning | reasoning enabled |

If Ollama does not report a capability, this plugin does not invent it from the model name.

Context Limits

Ollama model metadata can report very large architecture context windows, such as 131072, while runtime num_ctx may still be much smaller. To avoid overfilling local models, this plugin uses this priority:

1. Explicit plugin context option
2. num_ctx from /api/show parameters or Modelfile
3. model_info.*.context_length only when useModelInfoContext is true
4. Safe default 4096

Development

npm install
npm test
npm run typecheck
npm run build

License

MIT