Plugins

This page explains how plugins package reusable runtime behavior behind configuration.

Why Plugins Exist

Plugins let NeMo Relay install reusable runtime behavior from configuration instead of requiring every application or framework integration to register the same middleware and subscribers by hand.

Plugins package reusable runtime components.

Plugin Configuration Model

A plugin configuration document combines a schema version, an ordered set of components, and a policy for unsupported configuration.

Version

The version identifies the configuration format expected by the plugin system.

Components

Components identify the runtime behavior to activate. Each component declares its kind and configuration.

Policy

Policy controls how strictly the plugin system interprets unknown fields, unsupported values, or compatibility issues.

Two similarly named files serve different purposes:

Refer to Plugin Configuration Files for runtime configuration and Package Discoverable Plugins for the package manifest.

Component Lifecycle

Relay validates plugin configuration before it activates any component.

Validation

Validation checks whether the supplied config is structurally and semantically acceptable before initialization.

Initialization

Initialization activates the configured components and registers their runtime behavior.

Activation Reporting

Reporting provides structured diagnostics about what activated successfully and what did not.

Deactivation and Cleanup

Relay tracks registrations created during plugin initialization. Clearing active plugin configuration removes those registrations and runs component cleanup. Dynamic plugin teardown also stops managed workers and releases native activations after their callbacks are no longer registered. Applications should clear plugin configuration during graceful shutdown and test teardown. Refer to Register Plugin Behavior for the binding lifecycle.

Setup Failures

Plugin validation and initialization happen during setup. If configuration is invalid, a component kind is unavailable, or initialization fails, callers should treat setup as failed before relying on the new runtime behavior. Activation reports are the public way to inspect what validated or activated.

Runtime behavior after activation still belongs to the installed component. For example, an exporter can report delivery failures without changing tool or LLM execution semantics. Keep those component-specific failure rules in the component guide rather than redefining them in the plugin concept.

The following diagram shows how a plugin configuration flows through validation and the component lifecycle into installed runtime behavior:

Plugin Context

The plugin context is the API that a component uses to register its behavior. It connects plugin configuration to the active runtime.

What Plugins Can Register

Depending on the component, a plugin can register middleware, subscribers, and runtime helpers. These are the same execution surfaces available elsewhere in Relay, but the component context gives them shared configuration and ownership. This includes activation-owned conditional middleware guardrails that temporarily control the eligibility of an existing global registration.

This is what makes plugins a packaging mechanism rather than a separate runtime model. Plugins do not replace scopes, middleware, or subscribers. They install them.

Ownership and Scope

Plugin initialization is process-level. It is intended for runtime components that should remain active across requests in one process. It is not permanent as clearing or replacing the active plugin configuration removes plugin-owned runtime behavior.

Scope-local behavior still matters after plugin installation, but the plugin system itself is a global activation layer.

Plugins install runtime behavior; they do not create a separate execution model. Scopes still own parentage and cleanup, middleware still owns execution ordering, and events still own the canonical runtime record.

Plugin Delivery Models

All plugin components use the same validation and activation model. They differ in how their component kind becomes available to the host:

Built-In Plugin Components

The core runtime registers the observability, nemo_guardrails, and pricing components before lookup, validation, and initialization. The CLI and the Python and Node.js bindings also register adaptive and pii_redaction. Direct Rust applications must register Adaptive and PII Redaction from their component crates before validating or initializing either kind: nemo_relay_adaptive::plugin_component::register_adaptive_component() and nemo_relay_pii_redaction::component::register_pii_redaction_component(). Applications can still register custom plugins.

Adaptive

Adaptive is implemented as a built-in plugin component. It is not a separate runtime model. It uses the same plugin system as custom components.

This matters conceptually because adaptive behavior is configured and activated through the same component lifecycle as other plugins. Direct Rust applications follow this sequence:

1. Call nemo_relay_adaptive::plugin_component::register_adaptive_component().
2. Validate the config.
3. Initialize the plugin system.
4. Inspect the activation result if needed.

Detailed adaptive configuration belongs in Adaptive Configuration, Adaptive Cache Governor (ACG), and Adaptive Hints.

Observability

The core crate ships a built-in observability plugin component for Agent Trajectory Observability Format (ATOF), Agent Trajectory Interchange Format (ATIF), and typed OpenTelemetry exporters. Each exporter section is disabled unless its section sets enabled = true, and subscriber names are inferred from the plugin namespace instead of exposed in public config.

Detailed observability plugin configuration belongs in Observability Configuration.

NeMo Guardrails

The core crate also ships a built-in nemo_guardrails plugin component. The built-in integration is deprecated and scheduled for removal in NeMo Relay 0.9. There is no replacement in NeMo Relay 0.8; any replacement will target 0.9 or later. Do not use the built-in component for new deployments.

The current user-facing paths are the remote backend for Guardrails-service integration and the Python-backed local backend that runs nemoguardrails through a subprocess worker.

Detailed Guardrails plugin configuration belongs in NeMo Guardrails Configuration.

PII Redaction

The pii_redaction component sanitizes emitted observability payloads without changing real callback arguments or results. The CLI and primary language bindings register this component. Direct Rust applications must call nemo_relay_pii_redaction::component::register_pii_redaction_component() before they validate or initialize a PII Redaction component.

Configure actions, detectors, targets, and backend modes through PII Redaction Configuration.

Model Pricing

The core crate ships a built-in pricing component. It loads catalog sources that response codecs can use to annotate managed or manually observed LLM responses with cost estimates. Configure catalog sources through Model Pricing.

For plugins.toml discovery, precedence, merge, and gateway editing rules, refer to Plugin Configuration Files.

Discoverable Plugins

Discoverable plugins use the same component lifecycle, but the CLI reads a relay-plugin.toml manifest before it creates an internal component. The manifest identifies a Rust native shared library or a local grpc-v1 worker, declares compatibility and capabilities, and supplies integrity evidence for the artifact.

The operator keeps the manifest reference and component configuration in plugins.toml. Use nemo-relay plugins validate <plugin-id> to check the manifest, optional static schema, host policy, compatibility, and trust evidence before enabling or running a dynamic plugin. During startup, Relay loads the enabled adapter and then validates the synthesized component. Refer to Configure Discoverable Plugins for the operator workflow and Package Discoverable Plugins for the authoring workflow.

The operator lifecycle is explicit:

1. nemo-relay plugins add <manifest> records the manifest and prepares managed worker resources when required.
2. nemo-relay plugins inspect <plugin-id> reads the stored registration and resolved manifest without changing desired state.
3. nemo-relay plugins validate <plugin-id> checks compatibility, schema, policy, and trust evidence.
4. nemo-relay plugins enable <plugin-id> changes desired state; the plugin loads during the next host startup.
5. nemo-relay plugins disable <plugin-id> keeps the registration but prevents it from loading on the next startup.
6. nemo-relay plugins remove <plugin-id> deletes the registration and any Relay-managed worker environment after the running host has stopped.

These commands manage future host activation. Teardown of an already running host remains part of that host’s owned plugin cleanup.

Practical Guidance

• Reach for a plugin when runtime behavior is reusable across applications or integrations, so hosts install it from configuration instead of registering the same middleware and subscribers by hand.
• Validate configuration before initialization, and inspect the activation report to confirm what activated.
• Keep field-by-field settings in the guide for the component that owns them rather than restating them here.
• Clear plugin configuration during graceful shutdown and test teardown so plugin-owned registrations, managed workers, and native activations are released.
• Refer to the plugin authoring overview to choose whether a language-binding, native, or worker plugin is the right delivery model.