OpenRouter

{
  "publicModel": "search-reranker",
  "adapterKey": "openrouter",
  "upstreamModel": "cohere/rerank-4-fast",
  "credentials": { "apiKey": "sk-or-..." }
}

• Credentials: apiKey is required. baseUrl and custom headers are optional. The default base URL is https://openrouter.ai/api/v1.
• Operation: reranking only in this release. The adapter sends POST /rerank through the native openrouter_rerank transport.
• Provider routing: the public provider object is preserved exactly and makes only OpenRouter deployments eligible for that request.
• Model ids: use OpenRouter's complete ids, including creator and suffixes such as :free.
• Catalog: the committed catalog is generated from OpenRouter's model feed filtered by output_modalities=rerank. No expected model count is hard-coded.
• Text-only release: multimodal source capability is reported by the sync but withheld from the public profile. The NVIDIA vision reranker therefore accepts text documents only for now.
• Pricing: ambiguous zero token fields do not mark paid rerankers free. Reviewed search-unit overrides are used where available; otherwise OpenRouter's response cost is the accounting fallback.

Catalog workflow:

bun run --filter @boelabs/bifrost catalog:sync:openrouter
bun run --filter @boelabs/bifrost catalog:sync:openrouter:verify
# To update the committed snapshot after review:
bun run --filter @boelabs/bifrost catalog:sync:openrouter:write

The report calls out multimodal capability not yet enabled, ambiguous zero pricing, orphaned reviewed overrides, and paid models without a representable local cost.

What to read next

• Reranking — public contract, routing, validation, and examples.
• Model catalog — sync ownership and schema.
• Routing — retries, fallback, health, and price strategy.