feat: adding OpenRouter as an inference provider

[nv-kasikritc]

Problem Statement

When a developer onboards onto NemoClaw, NemoClaw still doesn't have an opinionated way to let developers to consume paid open-source model endpoints from our NCPs/ISPs—no continuation from free build.nvidia.com endpoints to paid endpoints.

We are making the inference routers (e.g. OpenRouter) our opinionated way to consume endpoints from NCPs.

This would offer an easy on ramp to paying for non-rate-limited models.

Proposed Design

During user onboarding, they should see OpenRouter as the second option during the inference configuration step.

Cloud models share the same list as NVIDIA Endpoints.

[3/8] Configuring inference provider
  ──────────────────────────────────────────────────

  Inference options:
    1) NVIDIA Endpoints
    2) OpenRouter
    3) OpenAI
    4) Other OpenAI-compatible endpoint
    5) Anthropic
    6) Other Anthropic-compatible endpoint
    7) Google Gemini
    8) Install Ollama (Linux)
    9) Model Router (experimental)

Choose [1]: 2

  ┌─────────────────────────────────────────────────────────────────┐
  │  OpenRouter API Key required                                    │
  │                                                                 │
  │  1. Go to https://openrouter.ai/workspaces/default/keys         │
  │  2. Sign in with your OpenRouter account                        │
  │  3. Click 'New Key' button                                      │
  │  4. Paste the key below (starts with sk-or-)                    │
  └─────────────────────────────────────────────────────────────────┘

  OpenRouter API Key: **********************************************************************

  Key staged for the OpenShell gateway. It is held in process memory only;
  onboarding registers it with the gateway and nothing is written to disk.


  Cloud models:
    1) Nemotron 3 Ultra 550B (nvidia/nemotron-3-ultra-550b-a55b)
    2) Nemotron 3 Super 120B (nvidia/nemotron-3-super-120b-a12b)
    3) Nemotron 3 Nano Omni 30B (nvidia/nemotron-3-nano-omni-30b-a3b-reasoning)
    4) GLM-5.2 (z-ai/glm-5.2)
    5) MiniMax M3 (minimaxai/minimax-m3)
    6) Kimi K2.6 (moonshotai/kimi-k2.6)
    7) DeepSeek V4 Pro (deepseek-ai/deepseek-v4-pro)
    8) Other...

  Choose model [1]:

API calls to OpenRouter should include these fields in the headers for app attribution.

"HTTP-Referer": "https://www.nvidia.com/nemoclaw/",
"X-OpenRouter-Title": "NVIDIA NemoClaw",

Alternatives Considered

No response

Category

Feature

Checklist

• I searched existing issues and this is not a duplicate
• This is a design proposal, not a "please build this" request

[wscurran]

@ericksoa @jyaunches was chatting with Kasikrit and Kris about this. This seems like the right direction for enterprise inference support and providing a path for non rate limited models. How big is this work? Is it something we can ship in Sprint 8 (next Sprint)?

[sandl99]

Confirmed with @nv-kasikritc @wscurran , We will bring OpenRouter as first class an inference provider for NemoClaw. I'll following up and add it. After triage, the onboarding is not big, we can delivery quickly.