A registry picks one of these adapters. Chat is always available. Embeddings, rerank, files, images, and audio only run when the adapter advertises them and the model slug can do that work.
Connect a card from Agent Gateway → Registry. How credentials and health checks work is on Registries. Clients still call one TrustGate path — chat, embeddings, images, files, audio, or rerank — and TrustGate adapts the wire format. MCP servers are a different registry type. See the MCP Gateway.Capability matrix
Code is the registry provider id (owned_by on
GET /v1/models, and the @provider/… prefix in
model resolution).
Text
Files, images, and audio
Yes means the adapter can forward that modality. A request still needs a
model that actually supports it — for example Bedrock embeddings only accept
Titan embed slugs, and OpenAI image calls need an image model.
Pinning a registry that cannot serve the path is a 400, not a failover. An
empty capable pool is a 503. See each endpoint page for the candidate pool.
What each capability maps to
Credentials
The credential on the registry is how TrustGate calls the provider. It is not the consumer key your applications send.
Use Custom / OpenAI-compatible when the upstream speaks Chat Completions
(and optionally
/v1/embeddings, images, or audio) but is not a first-class
card — vLLM, Ollama, LiteLLM, and similar.
Related
- Registries — connect, test, and price a backend
- Chat
- Model resolution
- Models