ModelCatalog.Client 0.2.0

ModelCatalog

A canonical catalog of LLM model metadata — pricing, context windows, capabilities — aggregated from LiteLLM, OpenRouter, and models.dev. Self-host or use the public PinkRoosterAI instance.

Public endpoint

https://models.pinkrooster.nl

Read-only, no authentication. Intended for low-volume use; if you need higher throughput or stronger SLAs, self-host (see below).

Quickstart

.NET client

dotnet add package ModelCatalog.Client

using ModelCatalog.Client;

services.AddModelCatalogClient(opts =>
{
    opts.BaseUrl = "https://models.pinkrooster.nl";
});

public sealed class MyService(IModelCatalogClient catalog)
{
    public async Task<decimal?> GetPriceAsync(string provider, string modelId, CancellationToken ct)
    {
        var info = await catalog.GetModelAsync(provider, modelId, ct);
        return info?.Pricing?.InputCostPerMillion;
    }
}

Raw HTTP

curl https://models.pinkrooster.nl/v1/models/openai/gpt-4o

API reference

Self-hosting

docker run -d \
  --name model-catalog \
  -p 8080:8080 \
  -v model-catalog-data:/app/data \
  ghcr.io/pinkroosterai/model-catalog:latest

To enable /v1/refresh, set at least one API key:

docker run -d \
  ... \
  -e ModelRegistry__ApiKeys__0__Name=admin \
  -e ModelRegistry__ApiKeys__0__Key="$(openssl rand -hex 32)" \
  ghcr.io/pinkroosterai/model-catalog:latest

The service fetches all three upstream catalogs on startup, then daily at 01:00 UTC.

How it works

1. Sources — Three pluggable normalizers (LiteLlmSource, OpenRouterSource, ModelsDevSource) fetch + canonicalize per-source data.
2. Aliases — A static map (Aliases/alias-map.json) reconciles cross-source naming differences (e.g. gpt-4o-2024-05-13 ↔ gpt-4o).
3. Merging — A per-field PriorityMerger picks the highest-trust value (default order: OpenRouter > LiteLLM > models.dev for prices; LiteLLM > others for context windows). Configurable via ModelRegistry:Merge in appsettings.
4. Snapshot — Merged catalog persisted to data/snapshot.json. Read endpoints serve this snapshot in-memory.
5. Refresh — Quartz cron + manual /v1/refresh. Fetch failures from one source don't break the snapshot — the others succeed and the failed source surfaces in /v1/meta.

Configuration

All configuration lives under the ModelRegistry section. Defaults are sensible; override via appsettings.json or ModelRegistry__* environment variables.

License

MIT © 2026 PinkRoosterAI