Glossary
This glossary defines the terms that the Fabric AI Gateway console and docs use. For how the main concepts fit together, see Key concepts.
| Term | Definition |
|---|---|
| Access mode | A setting on a connection or route: Open doesn’t require a grant, and Gated requires a grant for the caller. |
| Adaptive Model Router | A routing plugin that picks a model and reasoning effort for each LLM request. See Route requests with the Adaptive Model Router. |
| Alias | A model name with no /, defined on a route that routes by model name, that maps to one or more provider models. See Model aliases. |
| API key | An administrative key for managing your organization through the admin API. It’s separate from a gateway key. See Manage API keys. |
| Base path | The path prefix of a route, such as /anthropic. The route’s endpoints sit under it. |
| Base URL | The data plane host followed by a route’s base path. Clients point their SDK at it. |
| Circuit breaker | A check that stops sending traffic to an upstream after repeated failures, then lets a trial request through after a pause. See Limits. |
| Connection | The link between the gateway and an upstream: an LLM provider, MCP server, REST API, or agent. It holds the target URL. It’s never a network connection. The Services page and the admin API call the same thing a service. |
| Consumer | The identity of a caller, such as a person, team, or application. A gateway key or token identifies the consumer, and plugins, grants, and analytics use it. |
| Consumer group | A set of consumers that plugins and grants can target together. |
| Control plane | The part of Fabric AI Gateway where you configure routes, plugins, and access. Postman hosts it on every plan. It isn’t in the request path. |
| Credential | A secret that the gateway uses: a gateway key that identifies a consumer, or an upstream credential, such as a provider API key, that the gateway adds to requests. See Manage credentials. |
| Data plane | The part of Fabric AI Gateway that receives your applications’ requests, applies plugins, and calls upstreams. On the Free plan, Postman hosts it. On the Enterprise plan, it runs in your VPC. |
| Data plane host | The hostname that receives your applications’ requests, shown in the docs as GATEWAY_HOST. |
| Endpoint | A method and path that a route exposes, such as POST /v1/messages. See Supported endpoints. |
| Enterprise plan | A hybrid deployment: the control plane is hosted in Postman cloud, and the data plane runs in your VPC. See Choose a deployment model. |
| Fail closed, fail open | What happens when a plugin breaks. Fail closed refuses the request. Fail open skips the plugin and continues. |
| Failover | Moving a request to the next connection when one fails. The LLM Router configures it. See Fail over across providers with the LLM Router. |
| Free plan | The hosted plan: Postman hosts both the control plane and the data plane. |
| Gateway key | A key that a caller sends in the fg-api-key header to authenticate to a route. It also selects the caller’s namespace. See Authenticate and send a request. |
| Grant | Permission for a consumer to use a gated connection, or specific methods, paths, or MCP tools on it. |
| MCP | Model Context Protocol, which agents use to discover and call tools. The gateway serves MCP routes that combine one or more MCP servers. |
| Model filter | The models that a connection allows. A request for any other model gets 400. |
| Namespace | A space that separates teams’ connections, analytics, traces, and users within an organization. Every organization has a Default namespace. See Manage namespaces. |
| Organization | The account you sign in to. It holds your namespaces, members, and settings. |
| Plugin | A reusable behavior, such as authentication, rate limiting, or PII redaction, that you choose from the plugin catalog. See Control traffic with plugins. |
| Plugin instance | A configured copy of a plugin, applied at a scope and optionally limited by conditions. |
| Policy chain | The ordered list of plugin instances that runs for a request. Each one can inspect, allow, deny, or rewrite it. |
| Provider | A company or server that serves models, such as OpenAI or Anthropic. You connect one through a connection. |
| Route | The path that your applications send requests to. A route exposes the endpoints you choose under its base path and sends requests to its connections. |
| Scope | Where a plugin instance applies: globally, to a connection, to a route, or to a consumer or consumer group. |
| Session pinning | Keeping every turn of a conversation on the same model, so that the provider’s prompt cache keeps working. The Adaptive Model Router does this with a session header. |
| Tool Search | A plugin that replaces a route’s full MCP tool list with three tools that search for, describe, and run the others. |
| Trace | The record of one request through the gateway, including the plugins that ran and the upstream call. See Inspect traces. |
| Upstream | A system that the gateway calls on your behalf: an LLM provider, MCP server, REST API, or agent. |
| Upstream credential | A credential, such as a provider API key or your account at an MCP server, that the gateway adds to the request it sends upstream. |
| Vault | A credential backend that the gateway uses to look up upstream credentials. Sensitive values stay on the data plane. |
Next steps
- See how the concepts fit together in Key concepts.
- Read how a request moves through the gateway in Follow a request through the gateway.
- Find answers to common questions in the FAQ.

