Skip to content

providers.md

Latest commit

 

History

History
140 lines (92 loc) · 8.01 KB

providers.md

File metadata and controls

140 lines (92 loc) · 8.01 KB
Raw

Providers & Models

FindRomCover supports cloud providers, local servers, and custom endpoints. All of them are configured in Settings > AI Settings.... Test / Load Models discovers a provider's models at <Base URL>/models, so the endpoint must expose that path.

Provider comparison

Provider Default base URL Default model Key required Notes
OpenRouter https://openrouter.ai/api/v1 google/gemma-4-26b-a4b-it Yes Aggregator with hundreds of models; cheapest vision options
OpenAI https://api.openai.com/v1 gpt-4o-mini Yes Reliable JSON output; gpt-5-nano is a low-cost alternative
Anthropic https://api.anthropic.com/v1 claude-sonnet-4-5 Yes Strong visual reasoning; native Messages API
Gemini https://generativelanguage.googleapis.com/v1beta gemini-2.5-flash Yes Generous free tier; gemini-2.5-flash-lite is cheaper
GLM https://api.z.ai/api/paas/v4 glm-4.5v Yes Zhipu/Z.AI; use https://open.bigmodel.cn/api/paas/v4 for the China endpoint
Local http://localhost:11434/v1 qwen2.5vl:7b No Ollama or LM Studio; no cloud cost, slower on weak hardware
Custom (OpenAI-compatible) None — enter manually None — enter manually Optional Proxies, gateways, self-hosted servers
Custom (Anthropic-compatible) Falls back to https://api.anthropic.com/v1 when left empty None — enter manually Optional Endpoints speaking the Anthropic Messages API

OpenRouter

  1. Create an API key at openrouter.ai/keys.
  2. Select OpenRouter in AI Settings, paste the key, and click Test / Load Models.
  3. Pick a model. The default google/gemma-4-26b-a4b-it follows the cover/box-art instructions reliably; see Recommended Models for cheaper and more expensive alternatives.

OpenRouter exposes modality metadata, so the model picker knows exactly which models accept image input.

OpenAI

  1. Create an API key at platform.openai.com/api-keys.
  2. Select OpenAI, paste the key, and click Test / Load Models.
  3. Choose a vision-capable model such as gpt-4o-mini or gpt-5-nano.

The OpenAI adapter uses the shared OpenAI-compatible client. Reasoning models (o1, o3, o4, gpt-5*) are called without temperature and with max_completion_tokens; other models use temperature and max_tokens. All requests allow up to 4,096 output tokens, and JSON is parsed from the response text.

Anthropic

  1. Create an API key at console.anthropic.com.
  2. Select Anthropic, paste the key, and click Test / Load Models.
  3. Choose a model such as claude-sonnet-4-5 or claude-haiku-4-5 for lower cost.

Anthropic uses a native adapter implementing the Messages API, including image blocks.

Gemini

  1. Create an API key at aistudio.google.com/apikey.
  2. Select Gemini, paste the key, and click Test / Load Models.
  3. Choose a model such as gemini-2.5-flash or gemini-2.5-flash-lite.

Gemini uses a native adapter implementing generateContent with inline image data. The free tier is generous, which makes it a good starting point.

GLM

  1. Create an API key at z.ai.
  2. Select GLM, paste the key, and click Test / Load Models.
  3. Choose glm-4.5v or another vision-capable GLM model.

For the China endpoint, change the Base URL to https://open.bigmodel.cn/api/paas/v4.

Local (Ollama / LM Studio)

Run a local vision model and point FindRomCover at it. No API key is required.

Ollama

ollama pull qwen2.5vl:7b
ollama serve
  • Base URL: http://localhost:11434/v1
  • Model: qwen2.5vl:7b (or another vision model you have pulled)

LM Studio

  • Start the local server in LM Studio (default port 1234).
  • Base URL: http://localhost:1234/v1
  • Model: the identifier shown by LM Studio.

Local models are private and free to run, but they are slower and generally less accurate than cloud models. They are a good fit for small collections or offline setups.

Custom endpoints

Use a custom provider when you have a proxy, gateway, or self-hosted server:

Option API style Base URL When to use
Custom (OpenAI-compatible) /chat/completions Required — no default Anything that speaks the OpenAI API, including Azure OpenAI gateways and LiteLLM
Custom (Anthropic-compatible) /messages Optional — defaults to https://api.anthropic.com/v1 Endpoints that speak the Anthropic Messages API

Enter the API key, Base URL, and Model manually. The Base URL and Model are required for custom providers; the API key is optional at save time because some self-hosted servers ignore it, but gateways that enforce authentication need one.

Model discovery. FindRomCover lists models by appending /models to the configured base URL and accepts either the OpenAI (data) or Gemini (models) response shape. A custom gateway must expose that endpoint, or Test / Load Models fails. If it does not, type the model ID directly into the Model field and skip the test.

OpenCode Zen

OpenCode Zen is an OpenAI-compatible gateway from the OpenCode team. It currently offers MiMo-V2.5 Free, a multimodal model that accepts image input:

Setting Value
Provider Custom (OpenAI-compatible)
Base URL https://opencode.ai/zen/v1
API Key Zen key from opencode.ai/auth
Model mimo-v2.5-free

See Recommended Models for cost notes and caveats.

The model picker

Click Test / Load Models to:

  1. verify the API key and base URL by contacting the provider;
  2. download the list of available models;
  3. cache the list for the selected provider and base URL for 7 days.

The cache is stored in %LocalAppData%\FindRomCover\ai-models.json, keyed by provider and lowercased base URL. If a test fails, the error is shown and the previously cached list stays loaded. Because the cache key includes the base URL, changing the base URL selects a different (usually empty) entry — click Test / Load Models to populate it.

The picker:

  • shows vision-capable models only by default (based on provider modality metadata where available, and a name heuristic otherwise);
  • can be filtered with the filter box — type part of a model ID to narrow the list;
  • displays the number of loaded models;
  • falls back to the cached list instantly on later visits.

How FindRomCover talks to providers

All adapters share one abstraction with identical timeouts, JPEG downscaling, JSON parsing, and friendly error messages:

  • OpenAI, GLM, Local, and Custom (OpenAI-compatible) endpoints share one adapter using /chat/completions.
  • Anthropic and Custom (Anthropic-compatible) endpoints share the native Anthropic adapter using /messages.
  • Gemini uses its native adapter and calls generateContent with inline image data.

Every client requests up to 4,096 output tokens. OpenAI reasoning models (o1, o3, o4, gpt-5*) use max_completion_tokens and omit temperature; other models use temperature and max_tokens. FindRomCover always prompts for JSON and extracts the JSON object from the response text, so models without a native JSON mode still work.

The local pick prompt asks for the official cover or box art of the exact game and allows the model to answer -1 when no candidate is genuine cover art. The API-fallback prompt is slightly more lenient: it prefers a game cover or an in-game screenshot over a photo of the physical cartridge, cart, or disc, and accepts an in-game screenshot when no cover art exists. The prompt version participates in the verdict cache key, so prompt changes invalidate old cached answers.

Related pages