> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.nvidia.com/nemo/relay/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/nemo/relay/_mcp/server.

# Highlights

This page summarizes the notable capabilities in the current release
documentation set.

## NVIDIA NeMo Relay 0.4

NVIDIA NeMo Relay 0.4 adds host plugin installation, first-party PII redaction, local
Guardrails execution, pricing-aware observability, streaming raw-event export,
and stronger coding-agent trace fidelity.

### Compatibility Notes

* The Python-only NeMo Guardrails example plugin was removed. Applications
  should use the built-in `nemo_guardrails` component instead.
* Pricing estimates require an inline, file-backed, or embedding-provided
  pricing source. NeMo Relay does not ship provider price data by default.
* Local NeMo Guardrails mode requires Python 3.11 or newer and
  `nemoguardrails==0.22.0` in the runtime environment.

### CLI And Plugin Configuration

* Added `nemo-relay install`, `uninstall`, and `doctor --plugin` flows for
  Claude Code and Codex host plugins.
* Added generated local marketplace support so Claude Code and Codex can load
  NeMo Relay through native host plugin mechanisms.
* Layered code-driven plugin configuration over materialized global, project,
  and user plugin files while preserving documented file precedence.
* Added pricing source management commands for validating catalogs, adding
  file-backed sources, and resolving model pricing.

### Guardrails And Redaction

* Added a local Python-backed backend for the built-in `nemo_guardrails`
  component.
* Expanded Guardrails configuration docs for local and remote backend behavior,
  supported codecs, request defaults, tool boundaries, and streaming behavior.
* Added the first-party PII redaction plugin crate with deterministic local
  backend support and Python, Node.js, and WebAssembly helper surfaces.

### Observability

* Added ATOF streaming endpoints for HTTP POST, WebSocket, and long-lived
  NDJSON uploads.
* Added ATIF HTTP storage export alongside S3-compatible storage destinations.
* Added model-pricing lookup and cost layering for managed LLM responses.
* Propagated normalized LLM cost into ATIF step metrics, OpenInference
  attributes, and OpenTelemetry attributes.
* Set OpenTelemetry span status from NeMo Relay execution results.
* Preserved sanitized LLM request payloads in annotations and start-event
  output.

### Integrations

* Improved Hermes hook injection, routed-provider observability, wrapped ATIF
  fidelity, subagent lineage, and error-path export consistency.
* Added Codex observability contract coverage and suppressed noisy Claude Code
  lifecycle events.
* Added nested subagent session lineage and more consistent cost, provenance,
  replay, placeholder, and hook-only fallback exports for OpenClaw.
* Improved LangChain input serialization and fixed plugin context-manager
  deadlock behavior.
* Added flattened OpenInference LLM attributes for annotations and replay.
* Annotated Deep Agents model responses and refreshed LangGraph callback
  coverage.

### Documentation And Tooling

* Updated installation documentation for 0.4 packages and host plugin setup.
* Clarified Hermes integration paths and coding-agent plugin source manifests.
* Added NVSkills CI request workflow and regenerated NeMo Relay skill eval
  datasets.
* Updated dependency attribution files and isolated Rust test behavior for CI
  and SonarQube cleanup.

The complete changelog and release artifacts can be viewed on
[GitHub Releases](https://github.com/NVIDIA/NeMo-Relay/releases).