Skip to content

Non-interactive onboard fails with sk-* keys from Inference Hub #1380

Description

@rtruong-nvidia

Summary

Non-interactive onboarding with NEMOCLAW_PROVIDER=cloud fails with HTTP 401 when using sk-* API keys issued by NVIDIA Inference Hub (https://inference.nvidia.com). The NVIDIA Endpoints endpoint validation rejects the key.

Context

NVIDIA Inference Hub now issues sk-* keys exclusively. The older nvapi-* keys were generated via NGC (https://org.ngc.nvidia.com) and are no longer available from Inference Hub's Key Management UI.

Any NemoClaw code path that validates against the nvapi-* prefix or hits an endpoint that only accepts nvapi-* keys is now incompatible with the current key format.

Steps to Reproduce

  1. Get an API key from https://inference.nvidia.com → Key Management → Personal key (will be sk-* format)
  2. Run non-interactive onboard on a CPU Brev instance:
NVIDIA_API_KEY="sk-..." \
NEMOCLAW_NON_INTERACTIVE=1 \
NEMOCLAW_ACCEPT_THIRD_PARTY_SOFTWARE=1 \
NEMOCLAW_SANDBOX_NAME=my-assistant \
NEMOCLAW_PROVIDER=cloud \
NEMOCLAW_POLICY_MODE=suggested \
nemoclaw onboard
  1. Onboarding fails at step [3/7] Configuring inference (NIM):
[non-interactive] Provider: build
NVIDIA Endpoints endpoint validation failed.
Responses API: HTTP 401: Authentication failed | Chat Completions API: HTTP 401: Authentication failed

Expected Behavior

sk-* keys from Inference Hub should be accepted by the cloud provider, since that is now the only key format available from inference.nvidia.com.

Environment

  • NemoClaw: v0.0.3
  • OS: Ubuntu (Brev n2d-standard-4 CPU instance)
  • Node.js: v22.22.2
  • Docker: pre-installed on Brev
  • GPU: None (CPU-only instance, using cloud inference)

Workaround

None found. The nim provider requires a local GPU. The cloud provider rejects sk-* keys. Interactive onboard may allow manual configuration but the non-interactive path is blocked.

Suggestion

Update the cloud provider's endpoint validation to accept sk-* keys and route them to https://inference-api.nvidia.com (the endpoint that works with sk-* keys for OpenAI-compatible completions). This is already how OpenClaw handles sk-* keys successfully.

Metadata

Metadata

Assignees

Labels

area: installInstall, setup, prerequisites, or uninstall flowarea: onboardingOnboarding FSM, provider setup, sandbox launch, or first-run flowplatform: brevAffects Brev hosted development environmentsplatform: ubuntuAffects Ubuntu Linux environments

Type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions