How to Set Up NVIDIA NIM Custom Providers in OpenClaw (Kimi K2.5 Example)

Want to use NVIDIA NIM models like Kimi K2.5 in OpenClaw? This guide walks you through the exact configuration, common mistakes, and what actually controls web access.

Why NVIDIA NIM?

NVIDIA NIM (NVIDIA Inference Microservices) lets you access powerful models through a unified OpenAI-compatible API. Models like Kimi K2.5 from Moonshot AI, Llama variants, and others are available — often with generous free tiers.

OpenClaw supports NIM as a custom provider, but the setup has a few gotchas that trip people up.

The Correct Configuration

Here's a working config for Kimi K2.5 via NVIDIA NIM:

{
  "models": {
    "mode": "merge",
    "providers": {
      "nvidia": {
        "baseUrl": "https://integrate.api.nvidia.com/v1",
        "apiKey": "${NVIDIA_API_KEY}",
        "api": "openai-completions",
        "models": [
          {
            "id": "moonshotai/kimi-k2.5",
            "name": "Kimi K2.5",
            "reasoning": true,
            "input": ["text"],
            "contextWindow": 256000,
            "maxTokens": 8192
          }
        ]
      }
    }
  },
  "agents": {
    "defaults": {
      "model": {
        "primary": "nvidia/moonshotai/kimi-k2.5"
      }
    }
  },
  "tools": {
    "web": {
      "search": {
        "enabled": true,
        "provider": "brave",
        "apiKey": "${BRAVE_API_KEY}"
      },
      "fetch": { "enabled": true }
    }
  }
}

Set your API key in the environment:

export NVIDIA_API_KEY="nvapi-xxxx"

Common Mistakes (And How to Avoid Them)

1. Don't add auth: "api-key"

This field doesn't exist in OpenClaw's custom provider schema. Authentication is handled automatically via apiKey or auth profiles. Remove any auth field.

2. Be careful with input: ["text", "image"]

Only add "image" if you've confirmed the NVIDIA NIM endpoint actually supports vision. Most Kimi endpoints are text-only. If you declare image support incorrectly, OpenClaw might try to route image tasks to this model and fail.

3. Match the model ID exactly

OpenClaw passes your models[].id directly to NVIDIA's API as the model parameter. If NVIDIA expects "moonshotai/kimi-k2.5", that's exactly what you need — not "kimi-2.5" or "moonshot/kimi-k2.5".

"But My Model Can't Access the Web!"

This is the #1 misconception. In OpenClaw, web access is controlled by tools, not models.

Models don't "come with" web search. You need to configure the web_search tool separately:

{
  "tools": {
    "web": {
      "search": {
        "enabled": true,
        "provider": "brave",
        "apiKey": "${BRAVE_API_KEY}"
      },
      "fetch": {
        "enabled": true
      }
    }
  }
}

This applies whether you're using Kimi, Claude, Llama, or any other model. Without this config, no web search happens.

Get a free Brave API key: brave.com/search/api

Quick Verification

After adding your config, run:

openclaw models status   # Check if your model appears
openclaw doctor          # Verify overall health

You should see your NVIDIA model listed with status ✓ online.

Full Working Config (Copy-Paste Ready)

Here's everything together:

{
  "models": {
    "mode": "merge",
    "providers": {
      "nvidia": {
        "baseUrl": "https://integrate.api.nvidia.com/v1",
        "apiKey": "${NVIDIA_API_KEY}",
        "api": "openai-completions",
        "models": [
          {
            "id": "moonshotai/kimi-k2.5",
            "name": "Kimi K2.5",
            "reasoning": true,
            "input": ["text"],
            "contextWindow": 256000,
            "maxTokens": 8192
          }
        ]
      }
    }
  },
  "agents": {
    "defaults": {
      "model": {
        "primary": "nvidia/moonshotai/kimi-k2.5"
      }
    }
  },
  "tools": {
    "web": {
      "search": {
        "enabled": true,
        "provider": "brave",
        "apiKey": "${BRAVE_API_KEY}"
      },
      "fetch": { "enabled": true }
    }
  }
}

Other NVIDIA NIM Models

The same pattern works for other NIM models. Just update the id to match what NVIDIA expects:

• nvidia/llama-3.3-70b-instruct
• moonshotai/kimi-k2.5
• meta/llama-3.1-405b-instruct

Check NVIDIA's documentation for the exact model IDs available.

Based on a 370+ message Discord thread in #help. Props to @satyashetty for asking and @Krill for the detailed breakdown.

Related docs:

• Custom Providers
• Web Tools Setup