NVIDIA Models Plugin for Claude Code

A Claude Code plugin that makes 12 NVIDIA NIM-hosted LLMs (Llama 3.x, Kimi K2, DeepSeek V3/R1, Nemotron, Mixtral, Mistral Large, Qwen 2.5, Qwen Coder) available to Claude — and nudges Claude to use them proactively when the user wants a second opinion, an ensemble comparison, or a model better suited to a specific task.

What this packages

Component	What it does
MCP server (server/server.py)	Wraps NVIDIA NIM's OpenAI-compatible endpoint as three MCP tools: list_models, nvidia_chat, nvidia_compare.
Skill (skills/nvidia-delegate/)	Description-driven trigger that makes Claude consider routing work to NVIDIA models when delegation would help — without the user having to ask.
Slash commands (commands/)	Explicit one-key access: /nvidia-compare, /nvidia-list, /nvidia-chat.

The MCP server is the same hardened code shipped at nvidia-models-mcp v0.1.1 — bounded inputs, sanitized errors, 60 s timeout, no live network in tests.

Install

Prerequisites

• Claude Code (the CLI / desktop client that supports plugins)
• uv on PATH
• Python 3.10+
• An NVIDIA API key from build.nvidia.com (starts with nvapi-)

1. Export your API key

Windows (PowerShell, persistent):

setx NVIDIA_API_KEY "nvapi-your-key-here"

macOS / Linux (shell profile):

export NVIDIA_API_KEY="nvapi-your-key-here"

Restart Claude Code so it picks up the env var.

2. Add the plugin

Claude Code installs plugins from marketplaces in two steps. Run these inside Claude Code:

/plugin marketplace add NAJEMWEHBE/nvidia-models-plugin
/plugin install nvidia-models@nvidia-models-plugin

The first line registers this repo as a single-plugin marketplace (cloned to your local cache). The second installs the nvidia-models plugin from it.

After install: MCP server is wired automatically, the skill loads at session start, and the three slash commands (/nvidia-compare, /nvidia-list, /nvidia-chat) become available. No restart needed — Claude Code hot-reloads plugins.

How the auto-routing works

Claude reads every skill's description field at session start. The nvidia-delegate skill description lists trigger scenarios ("second opinion", "ensemble", "what would model X say", explicit NVIDIA model names, etc.). When the user's request matches, Claude opens the skill, reads the routing table inside, and calls the appropriate MCP tool. You never have to say "use the MCP" — just describe what you want.

Examples that should auto-trigger:

"Get a second opinion from another LLM on this refactor." "Ask three models to draft a one-paragraph summary and compare them." "Use qwen-coder to write the test, then have deepseek-r1 critique it." "What would Kimi K2 say about this?"

If you want explicit control, use the slash commands instead.

Slash commands

Command	Purpose
/nvidia-compare <prompt>	Fan the prompt to 3 default models; synthesize agreements + divergences.
/nvidia-list	Show the 12 short keys → full NIM paths.
/nvidia-chat <short-key> <prompt>	Single-model call.

Models exposed

Catalog verified live on NIM as of 2026-05-12.

Short key	NIM path	Notes
llama-4-maverick	meta/llama-4-maverick-17b-128e-instruct	Meta flagship; 128-expert MoE
llama-3.3-70b	meta/llama-3.3-70b-instruct	Generalist
llama-3.1-70b	meta/llama-3.1-70b-instruct	Cheap generalist
deepseek-v4-pro	deepseek-ai/deepseek-v4-pro	Premium reasoning MoE
deepseek-v4-flash	deepseek-ai/deepseek-v4-flash	Fast reasoning MoE
kimi-k2.6	moonshotai/kimi-k2.6	Moonshot Kimi
qwen3-next-80b	qwen/qwen3-next-80b-a3b-instruct	Qwen MoE
nemotron-super-49b	nvidia/llama-3.3-nemotron-super-49b-v1	NVIDIA Nemotron flagship
nemotron-mini-4b	nvidia/nemotron-mini-4b-instruct	Tiny NVIDIA
mixtral-8x22b	mistralai/mixtral-8x22b-instruct-v0.1	Large Mistral MoE
mixtral-8x7b	mistralai/mixtral-8x7b-instruct-v0.1	Small Mistral MoE

Security

• NVIDIA_API_KEY is read from the environment at MCP server startup and never persisted or logged.
• All tool arguments are bounded (prompt ≤ 32 KB, system ≤ 8 KB, max_tokens ≤ 8192, temperature ∈ [0, 2]).
• Upstream error messages are sanitized — nvapi-…, sk-…, and Authorization/x-api-key substrings are replaced with [REDACTED] before reaching Claude.
• The MCP command is uv (PATH lookup). Verify your uv binary (uv --version) is the official astral build — a PATH-shadowed uv could exfiltrate the API key.
• Full threat model in the MCP repo's SECURITY.md.

Never paste your API key into a chat session. If you do, rotate it at build.nvidia.com.

Cost

NVIDIA NIM has a free monthly credit. After that, per-token pricing varies by model — check build.nvidia.com for current quotas.

Relationship to nvidia-models-mcp

• nvidia-models-mcp — the bare MCP server. Use directly if you're on Claude Desktop (drag-drop DXT), Codex CLI, or any non-Claude-Code MCP host.
• nvidia-models-plugin (this repo) — the Claude Code experience: bundles the same server plus skill + slash commands. The skill is what makes Claude proactively reach for these models.

Bug reports, model additions, and contributions welcome on either repo.

License

MIT