Skip to navigation

Control traffic with plugins

Plugins are where you configure how Fabric AI Gateway handles traffic. They can authenticate callers, protect sensitive data, transform requests and responses, observe LLM usage, and control traffic across LLM, MCP, and HTTP API routes.

Some behavior is provided by system plugins. For example, creating a connection with Gateway auth set to Gateway key adds Key Auth to the new route. Other plugins are configured as plugin instances that you attach to a scope.

Start with a capability

  • Authenticate callers: Key Auth requires a gateway key on a route. Fabric adds it when you create a connection with Gateway auth set to Gateway key. See Authenticate requests with Key Auth.
  • Protect sensitive data: PII Redaction detects and redacts configured personal data in LLM and MCP traffic.
  • Transform requests and responses: Transformer can rewrite headers, query parameters, bodies, and cookies on supported routes, including LLM, MCP, and imported HTTP APIs.
  • Observe LLM traffic: Use Analytics and traces to review usage, cost, latency, errors, and individual requests.
  • Control cost: Set spending budgets and token-based limits, and use adaptive model routing. See Control spend with budgets.

Plugins run in an ordered policy chain, where they can inspect, allow, deny, or rewrite traffic at different points in the request lifecycle. The catalog below lists built-in plugins by category and the following sections show how to configure them.

Browse plugins by category

The catalog groups plugins into the following categories: Auth, Governance, Observability, Routing, Traffic, and Transform.

CategoryPluginWhat it does
AuthJWT AuthAuthenticates bearer JWTs and projects the verified identity into request context.
AuthKey AuthAuthenticates API keys and projects the matched consumer identity into request context. Fabric adds it to a route when you create a connection with Gateway auth set to Gateway key.
GovernanceAccess ControlEnforces configured access rules using the resolved caller, route, connection, and request context.
GovernanceAgent ApprovalRequires an approval token or IAM scope before an agent action is allowed.
GovernanceAgent Scope GateChecks agent and OAuth scopes against the route or MCP tool requirements.
GovernanceAgent Static DenyBlocks agent traffic that matches configured static deny rules.
GovernanceDeny AllRejects every request at an attached scope.
GovernanceKillswitchImmediately blocks traffic for configured tenant, connection, route, or MCP tool scopes.
GovernanceMCP HTTP bridgeServes a REST connection on an MCP route as MCP tools compiled from its OpenAPI spec.
GovernancePII RedactionDetects and redacts configured PII in LLM and MCP traffic. Attach it to routes that handle sensitive data, such as Social Security or phone numbers.
GovernanceProvider PassthroughPasses a route request directly to its configured provider.
GovernanceTool SearchAdds semantic discovery meta-tools for an MCP tool catalog. Instead of showing a model every tool definition, it shows three tools (search, get, and execute) that find the others, so the model sees only what it needs, when it needs it.
GovernanceValidatorValidates configured request and response fields through policy rules.
GovernanceLLM budgetEnforces UTC calendar spending limits using synchronized provider pricing, Redis counters, and control-plane checkpoints.
GovernanceLLM ObservabilityRecords token usage, the response model, and the finish reason for each LLM request.
GovernanceAgent AuditEmits an audit record for each completed or failed agent action.
ObservabilityAccess LogRecords completed gateway requests for operational auditing and diagnostics.
ObservabilityToken CounterCounts LLM token usage after a request completes.
RoutingLLM Model RouterRoutes a <provider>/<model> request to the connections that serve that model, in order, until one answers.
RoutingAdaptive Model RouterSelects an LLM model and reasoning effort per request using configurable heuristics, classifiers, or structured decision signals.
RoutingLLM RouterChooses which of the route’s connections handles each LLM request, and fails over when one fails.
TrafficLLM Rate LimitEnforces LLM request and token-based limits.
TrafficRate LimitEnforces request-rate limits for the attached scope.
TransformSecret InjectorResolves configured secrets and injects them into outbound request headers or query parameters.
TransformTransformerTransforms configured request and response fields through policy obligations.

The Transformer and Validator plugins are the most versatile. Both can run in the request and response phases and can read nearly every available field. Transformer can change headers, query parameters, body, status, and cookies. Validator only adds a marker header. They can do so only when scoped globally, to a connection, or to a route.

Add a plugin instance

A plugin in the catalog doesn’t do anything on its own. To apply it to traffic, create a plugin instance that attaches the plugin to a scope and configures it.

To add a plugin instance, do the following:

  1. In the sidebar, select Plugins.
  2. On the All tab, find the plugin you want, such as Transformer or Rate Limit. Use Search plugins or the filters to narrow the list.
  3. Next to the plugin, click Use. The New plugin instance form opens.
  4. Enter a Name for the instance. Names must be unique. Reusing an existing name replaces that instance. You can’t change the name after you create the instance.
  5. Select a Scope kind, such as global or route. Some plugins can be scoped only to a route. If you choose a scope narrower than global, also select one or more Scope name values, such as specific routes. Fabric creates one instance for each.
  6. Under Configuration, configure the plugin using its dedicated form, or click Edit as JSON to provide raw configuration. Edit as JSON isn’t available for Rate Limit, Passthrough, or Adaptive Model Router.
  7. Optionally, expand Match conditions and status and add Match conditions to limit when the instance runs:
    • Path prefix
    • Methods
    • Only run for consumer
    • Only run for consumer group
  8. Under Status, confirm that Enabled is on. Fabric saves disabled instances but skips them at request time, so you can stage a configuration without activating it.
  9. Click Save.

To review the plugins that you’ve already set up, select the Configured tab.

Example: Rewrite a header with the Transformer plugin

The following example adds a transform that sets an X-Api-Key header on every matching request.

  1. Add a plugin instance and select Transformer as the plugin.
  2. Under Transforms, set the following values:
    • Phase: Request
    • Operation: set
    • Target: Header
    • Name: X-Api-Key
    • Value: prod
  3. Review the Compiled policy (Rego) panel. This panel is a read-only preview of the OPA policy that the control plane compiles from your transform and ships to the data plane’s sidecar.
  4. Click Save.

Example: Rate limit an MCP route

The following example limits traffic to a specific MCP route.

  1. Add a plugin instance and select Rate Limit as the plugin.
  2. Set Scope kind to route and Scope name to the route you want to limit, for example linear-mcp.
  3. Under Configuration, set Count by to one of the following:
    • Consumer: One counter per authenticated caller. Requires authentication on the route.
    • Credential: One counter per API key. Fabric fills this counter only when Key Auth is in use.
    • Global: One counter for every request. This is the only option that works without authentication.
  4. Set Limit, the number of requests allowed per window. This field is required.
  5. Set Window (seconds). The default is 60.
  6. Click Save.

To see how a plugin’s execution appears in request traces, see Inspect traces.

Guides for specific plugins

GuideUse it to
Authenticate requests with Key AuthRequire a gateway key on a route.
Route LLM requests across models and providersPick a model for each request, and fail over between providers.

Next steps