What you can do
Call any connected model
Point an OpenAI or Anthropic SDK at Atlan and call it.
Move existing code over
Change a base URL, key, and model name. Nothing else.
Connect your providers
Add provider credentials once and choose which models agents call.
Fall back across providers
Spread calls over deployments, retry, and keep them in-region.
Reuse identical answers
Serve a repeated request from cache without calling the provider.
See who spends what
Review usage by model, person, and agent, down to each call.
Get started
Connect client to Model Gateway
Create an agent key, list the models you can call, and make your first request.
Key concepts
Common tasks
Availability and limits
- Only organization admins add providers and turn models on or off. Everyone else calls the models that are on.
- A model with no price lands off, and can’t be turned on until it has rates.
- The gateway serves selected OpenAI and Anthropic operations. Files, batches, assistants, vector stores, fine-tuning, and Anthropic token counting return
404. - The Responses API is stateless: each request carries its whole conversation.
- Calls an agent makes with its own key appear in usage reports for organization admins only.
Reference
- Model API compatibility: endpoints, headers, model names, and request limits.
- Supported model providers: provider types, credentials, and discovery.
- List available models: the qualified names your client can call.
- OpenAI-compatible chat completion: request and response schemas.
Troubleshooting
See Model Gateway errors for refused calls and models missing fromGET /models.