The connections.
Filter by what your application needs, then open a provider for its endpoint, credential, and fallback detail.
- Providers
- 11
- Wire protocols
- 3
- Bring your own key
- 9
- Built-in presets
- 9
Showing 11 of 11 providers
- ProtocolOpenAIEndpointsModels16 listedCredentialsBYOKSetupRegistry routeUpstream statusOperationalChat Completions: Operational · Responses: OperationalOpenAIDirect model provider
- ProtocolAnthropicEndpointsModelsNot publishedCredentialsBYOKSetupRegistry routeUpstream statusNot publishedAnthropicDirect model provider
- ProtocolOpenAI-compatibleEndpointsModelsNot publishedCredentialsBYOKSetupBuilt-in presetUpstream statusNot publishedGeminiDirect model provider
- ProtocolOpenAI-compatibleEndpointsModelsNot publishedCredentialsBYOKSetupBuilt-in presetUpstream statusNot publishedMistral AIDirect model provider
- ProtocolOpenAI-compatibleEndpointsModelsNot publishedCredentialsBYOKSetupBuilt-in presetUpstream statusNot publishedDeepSeekDirect model provider
- ProtocolOpenAI-compatibleEndpointsModelsNot publishedCredentialsBYOKSetupBuilt-in presetUpstream statusNot publishedGroqInference provider
- ProtocolOpenAI-compatibleEndpointsModelsNot publishedCredentialsBYOKSetupBuilt-in presetUpstream statusNot publishedTogether AIInference provider
- ProtocolOpenAI-compatibleEndpointsModelsNot publishedCredentialsBYOKSetupBuilt-in presetUpstream statusNot publishedFireworks AIInference provider
- ProtocolOpenAI-compatibleEndpointsModelsNot publishedCredentialsBYOKSetupBuilt-in presetUpstream statusNot publishedOpenRouterModel routing platform
- ProtocolOpenAI-compatibleEndpointsModelsNot publishedCredentialsNo upstream keySetupBuilt-in presetUpstream statusNot publishedvLLMSelf-hosted inference runtime
- ProtocolOpenAI-compatibleEndpointsModelsNot publishedCredentialsNo upstream keySetupBuilt-in presetUpstream statusNot publishedOllamaLocal model runtime
How to read the table.
All 11 rows connect the same way. They differ on four columns, and which column matters depends on what your application calls today.
Protocol is the request family upstream.
Each provider's own API accepts one request family, and Relays speaks it on the far side of the edge. Your client keeps the contract it already has, so this column tells you what Relays talks to, not what you have to write.
- OpenAI-compatible
- 9
- Anthropic
- 1
- OpenAI
- 1
Endpoints decide who can cover for whom.
This is the set of Relays edge endpoints a connection serves. A fallback candidate has to declare the same edge endpoint as the route it backs, so two providers can only cover for each other where their endpoint sets overlap.
- Chat Completions
- 8
- Chat Completions · Responses
- 2
- Messages
- 1
Credentials say whose key runs the call.
9 connections take a provider key you already hold. The other 2 declare no upstream key by default, because they run against a runtime you host. Each provider page names the exact credential it expects.
- BYOK
- 9
- No upstream key
- 2
Setup is how much you declare.
9 providers arrive as a built-in preset with the base URL, protocol, and endpoint set already declared, so adding a key is the whole setup. The other 2 are registry routes you write yourself, which is also what you write when your endpoint or credential differs from the default.
- Built-in preset
- 9
- Registry route
- 2
Model catalogues, deprecations, and upstream status are shown only where the provider publishes them. Where nothing is published, the directory says so.

Bring the provider you already use.
Tell us the endpoint your application calls today and we will show you the route that replaces it.
