OpenCode AutoDiscovery Plugin

Auto-discovers models from OpenAI-compatible endpoints and populates opencode's model list. Works with llama.cpp, Ollama, LM Studio, and anything else that uses the /v1/models API.

Why

opencode doesn't auto-discover models from local endpoints. You have to manually define each model in your config. This plugin fixes that by hitting your endpoint's /models endpoint and loading whatever's available.

Features

• Discovers models from your OpenAI-compatible endpoint, respecting any you've already configured manually
• Beautifies ugly model IDs into readable names (qwen2-5-7b-instruct -> Qwen 2.5 7B Instruct)
• Detects context and output limits from standard fields (context_length, max_model_len, max_completion_tokens, etc.)
• Falls back to min(context/4, 32000) for output limits when undetectable
• Supports auth via apiKey in provider options
• Maps llama-router's API addon report: reasoning effort variants, the advertised default effort, a thinking-off variant, supported_parameters, and full pricing
• Works with any @ai-sdk/openai-compatible provider in your config

Installation

From git

Add to your opencode.json:

{
  "plugin": [
    "HarutoHiroki/OpenCode-AutoDiscovery#main"
  ]
}

From local path

{
  "plugin": [
    "/path/to/OpenCode-AutoDiscovery"
  ]
}

Configuration

No config needed. Just make sure your OpenAI-compatible provider is defined in your opencode.json:

{
  "provider": {
    "my-local-llm": {
      "npm": "@ai-sdk/openai-compatible",
      "options": {
        "baseURL": "http://localhost:8080/v1"
      }
    }
  }
}

For endpoints that require auth, add apiKey:

{
  "provider": {
    "my-gateway": {
      "npm": "@ai-sdk/openai-compatible",
      "options": {
        "baseURL": "https://gateway.example.com/v1",
        "apiKey": "your-api-key"
      }
    }
  }
}

The plugin will automatically discover models for any provider using @ai-sdk/openai-compatible. If you've already manually configured some models on a provider, only the remaining ones are discovered.

How it works

On opencode startup, the plugin:

1. Scans your providers for any using @ai-sdk/openai-compatible
2. Hits each provider's /v1/models endpoint (with Bearer auth if apiKey is set)
3. Reads context/output limits from standard response fields (context_length, max_model_len, max_completion_tokens, max_output_tokens, meta.n_ctx)
4. Skips models you've already configured manually
5. Beautifies model IDs into human-readable names
6. Registers discovered models with opencode

llama-router API addon

llama-router rewrites its /v1/models report through an API addon (src/api_addon.py) into an OpenRouter-style shape. The plugin maps that report onto opencode's model config:

• reasoning_effort.levels -> one variant per level, each sending reasoning_effort: <level> on the wire
• reasoning_effort.default -> the model's default reasoningEffort, so "no variant" matches the router's advertised default
• reasoning_effort.disable -> a none variant that turns thinking off: none sends reasoning_effort: "none", lowest sends the lowest level, qwen sends chat_template_kwargs: { "enable_thinking": false }
• supported_parameters -> tool_call from the presence of tools
• pricing -> cost, including cache_read and cache_write
• architecture.input_modalities / output_modalities -> input/output modalities
• context_length / max_output_tokens -> limit.context / limit.output

Models with a reasoning surface also get reasoning: true so opencode advertises the capability.

Model name beautification

The plugin has built-in knowledge of common model prefixes and tags:

• qwen2-5-7b-instruct -> Qwen 2.5 7B Instruct
• gemma-3-27b-it -> Gemma 3 27B Instruct
• glm-4-9b-chat -> GLM 4 9B Chat
• kimi-k2-0905-preview -> Kimi K2 0905 Preview

Unknown parts are capitalized and joined with spaces.

License

MIT, optionally credit this if you implement the code in your codebase, would be appreciated.