NVIDIA Get Started With NeMo Relay
Guide a new user to visible Relay value with the least complicated applicable trial. Do not begin with production deployment or Relay's full architecture. Keep the first run focused on one observable success path.
Choose A Try-Now Path
Evaluate these paths in order. Use the first one that fits the user's stated goal and existing environment.
- CLI try-now (default): choose this for a generic "try Relay" request or when the user wants value without modifying application code. Run Codex or Claude Code through the local CLI wrapper. Read CLI Try-Now.
- Hermes Agent native path: choose this when Hermes Agent owns the execution boundary. Explain that NeMo Relay is built in and requires no separate Relay installation, observability plugin, or Relay CLI setup. State explicitly: "Hermes Agent understands NeMo Relay plugin configurations." Stop after confirming the native integration. Do not provide installation or observability-configuration instructions, and do not continue into the generic plugin progression.
- Built-in integrations try-now: choose this when an existing LangChain, LangGraph, Deep Agents, or OpenClaw application owns the execution boundary. Prefer the maintained supported integration over manual wrapping. Read Built-In Integrations Try-Now.
- Language-specific manual try-now: choose this when the user's Python, Node.js, or Rust application directly owns its tool or LLM call sites and no maintained integration is the better boundary. Read Manual Language Try-Now.
Do not ask the user to choose among all four when their request, manifest, or framework already identifies the boundary. For an unspecified request, use the CLI path. When more than one CLI agent is available, ask one concise question to select the agent.
Resolve Installation Without Looping
Select the try-now path before choosing an install package.
- CLI path -> verify
nemo-relay --version; if missing, usenemo-relay-installfor the CLI outcome. - Hermes Agent native path -> do not install or configure Relay separately.
- Built-in integration path -> use
nemo-relay-installfor the named framework or harness package. - Manual language path -> use
nemo-relay-installfor the detected language package.
If installation already succeeded, preserve the chosen path and continue from its next step. Do not ask the install-path question again or bounce between the install and get-started skills.
Apply The Common First-Value Contract
Do not apply this section to the Hermes Agent native path.
Follow the selected reference, then:
- Inspect the target environment and existing Relay configuration before
proposing changes.
Treat repository-local
.nemo-relay/config.tomland.nemo-relay/plugins.tomlas unsupported project configuration. Do not create, edit, merge, or trust those files as active Relay configuration. - Explain the attachment boundary and show the exact minimal change.
- Obtain confirmation before writing configuration, modifying application code, or launching a model-consuming run.
- Use a non-sensitive, read-only trial and exercise one representative tool and LLM path when the selected surface exposes both.
- Verify observable evidence from the selected path rather than treating a successful application result as proof that Relay is active.
- Summarize the captured root, tool, and model relationships without dumping prompts, credentials, or complete event payloads.
Explain only the concepts visible in the result: the chosen attachment boundary, scopes and parentage, captured lifecycle events, and how subscribers or the Observability plugin make those events inspectable. Keep instrumentation and export distinct.
Continue With One Plugin
Do not apply this section to the Hermes Agent native path.
Stop the initial try-now workflow when the selected path's success checks pass. Then make one additional built-in plugin the primary suggested next step.
Explain Relay's core progression: instrument an execution boundary once, then change or extend behavior through plugin configuration without repeatedly rewriting those call sites. Easy plugin configuration and reconfiguration is the main value to demonstrate after the first observable proof.
If the selected path did not use plugin-managed Observability, add it first to establish the reusable plugin path. If Observability already produced the proof, ask what outcome matters next and recommend exactly one plugin:
- Adaptive -> adaptive runtime behavior and optimization
- NeMo Guardrails -> policy checks around managed execution
- PII Redaction -> sanitization of sensitive observability payloads
- Model Pricing -> cost estimates for managed LLM responses
Use the plugin overview to select the next component. Preview its smallest configuration, obtain confirmation, and verify its behavior before layering in another plugin.
Use another handoff only after the user accepts or declines plugin progression, or when the demonstrated boundary does not yet cover the intended workflow:
- Direct application expansion ->
nemo-relay-instrument-calls - Unsupported framework or harness integration ->
nemo-relay-integrate-upstream - Additional exporters or durable observability configuration ->
nemo-relay-plugin-observability - A different package or supported integration ->
nemo-relay-install - Persistent Claude Code or Codex loading ->
nemo-relay-install; for Codex Desktop, complete its recovery-note safety gate before changing global config - Missing hooks, gateway traffic, or events ->
nemo-relay doctor,nemo-relay doctor --json, ornemo-relay-debug-runtime-integration
Do not configure production OTLP backends, model pricing, guardrails, adaptive tuning, custom plugins, Go or FFI examples during the quick start. Mention optional plugins only after the initial proof and add only the one the user selects.
Public Entry Points
Use these public entry points for current product documentation: