Frequently asked questions
This page answers common questions about Fabric AI Gateway. For problems with a specific error, see Troubleshooting.
Sending requests
Do I send my provider's API key to the gateway?
No. You send a gateway key in the fg-api-key header. The gateway removes it and adds the provider’s credential from the connection before it calls the provider. For details, see Authenticate and send a request.
Can I use the OpenAI or Anthropic SDK?
Yes. Point the SDK’s base URL at the route’s base URL, and add the fg-api-key header with your gateway key. Request and response bodies are the provider’s own, so the SDK works as usual. For examples, see Use the OpenAI SDK and Use the Anthropic SDK.
Does the gateway translate between OpenAI and Anthropic formats?
No. The gateway sends an OpenAI-format request only to a provider that offers that OpenAI endpoint, and an Anthropic Messages request only to an Anthropic-format endpoint. A request to an endpoint that no provider on the route offers gets 404 unknown LLM endpoint. For which providers offer which endpoints, see Supported endpoints.
Does streaming work through the gateway?
Yes. Set "stream": true as you would with the provider. The gateway passes each server-sent event to you as soon as it’s complete.
How do I know which model answered my request?
Check the model field in the response. On a route that routes by model name, the x-fabric-model response header also names the provider and model that answered. In Traces, you can see every attempt the gateway made. For details, see Check which model answered.
How big can a request be?
The gateway reads up to 10 MiB of an LLM request body and up to 8 MiB of an MCP request body. Larger MCP requests get 413. For every limit, see Limits and defaults.
Can I call the gateway from a browser?
Not directly. The gateway doesn’t handle CORS, so browsers block the responses. WebSocket connections also don’t work through the gateway. Call it from a server or a backend for your web app.
Errors and retries
How do I tell a gateway error from a provider error?
When a provider answers with an error, the gateway relays it unchanged, so you get the provider’s own error body. Gateway errors use the bodies listed in Gateway error codes. To be sure, find the request by its X-Request-Id in Traces and check whether the gateway called the provider.
Does the gateway retry failed requests?
Not by default, with one exception: a request whose upstream credential is signed, such as AWS, gets one more try with fresh credentials after a 401 or 403. To retry or fail over to another provider, add the LLM Router to the route. It fails over on 429 and 503 by default, and also on 500, 502, and 504 for LLM routes. The gateway never retries after it starts sending a response. For details, see Fail over across providers with the LLM Router.
Why don't streamed OpenAI chat requests show token usage?
OpenAI includes usage in a chat stream only when the request sets stream_options.include_usage to true. The gateway doesn’t add that option for you. Set it in your request to see token counts in analytics.
MCP
Why do my MCP tool names have a prefix?
On a route with several MCP connections, the gateway combines their tools into one list. To keep names unique, each tool name starts with a prefix for its connection, in the form <prefix>__<tool>. On a route with one connection, names are unchanged.
Why do I see a connect_<service> tool instead of the server's tools?
Why do I see a connect_<service> tool instead of the server's tools?
The MCP server needs an account of your own, and you haven’t connected it yet. Call the tool, or open the link it returns, to connect your account. For details, see Connect your accounts.
Deployment
Can I run the whole gateway myself?
No. On the Free plan, Postman hosts both the control plane and the data plane. On the Enterprise plan, the control plane is hosted in Postman cloud and the data plane runs in your VPC. There is no option to run the control plane yourself. For details, see Choose a deployment model.
What happens to traffic if the control plane is unreachable?
Your data plane keeps serving traffic with the latest configuration it received. Configuration changes that you make in the meantime reach it when the connection comes back. For details, see How Fabric works.
What do the health check endpoints return?
/healthz returns 200 with {"status":"ok"} while the process is up. /readyz returns 200 with {"status":"ready"} while the process can take requests, even before the data plane has received its configuration. It answers 503 only on a data plane configured with an OPA policy bundle that hasn’t loaded yet. Both answer on any hostname and need no credential. For details, see Gateway API reference.
Next steps
- Look up gateway terms in the Glossary.
- Fix common errors with Troubleshooting.
- Send your first request by following Quickstart: Send your first request on the hosted gateway.

