nvidia-delegate

Delegate to NVIDIA-hosted LLMs

Three MCP tools wired by this plugin:

• mcp__nvidia-models__list_models — short keys → full NIM paths.
• mcp__nvidia-models__nvidia_chat(model, prompt, system?, max_tokens?, temperature?) — single-model call.
• mcp__nvidia-models__nvidia_compare(prompt, models?, system?, max_tokens?) — fan the same prompt across N models in one call. Default trio: llama-4-maverick, deepseek-v4-pro, qwen3-next-80b (three different vendors / architectures).

When to invoke

User intent	Tool	Suggested model
"Second opinion" / "what would another LLM say"	nvidia_compare	default 3
"Compare answers from X, Y, Z"	nvidia_compare	user-specified list
"Quick draft" / "cheap inference"	nvidia_chat	llama-3.1-70b or mixtral-8x7b
Tiny / very fast	nvidia_chat	nemotron-mini-4b
Reasoning-heavy / chain-of-thought	nvidia_chat	deepseek-v4-pro or deepseek-v4-flash
Long-context generalist	nvidia_chat	llama-4-maverick
NVIDIA-tuned generalist	nvidia_chat	nemotron-super-49b
Qwen-family preference	nvidia_chat	qwen3-next-80b
"Ensemble" / "voting" / "multi-agent"	nvidia_compare	3–5 generalists
Explicit "use "	nvidia_chat	the one named

Defaults to pick when the user is vague

• General Q&A → llama-4-maverick (or llama-3.3-70b for cheaper)
• Reasoning → deepseek-v4-pro (premium) or deepseek-v4-flash (fast)
• Cheap throwaway draft → nemotron-mini-4b
• Comparison → nvidia_compare (default trio: llama-4-maverick, deepseek-v4-pro, qwen3-next-80b)

After the call

• Don't dump raw model output unless the user asked for it.
• For nvidia_compare: summarize agreements + disagreements, then surface the most useful answer.
• For nvidia_chat: integrate the response into Claude's own answer; cite which model said what when it matters.

When NOT to invoke

• The user is in a tight feedback loop debugging the current Claude conversation — don't hand off mid-flow.
• A trivial factual lookup that doesn't benefit from a second model.
• The user explicitly said "you answer this" or "don't delegate".
• The prompt contains secrets or PII the user wouldn't want sent to a third party — surface that risk first.

Failure modes

• NVIDIA_API_KEY not set → tell the user: setx NVIDIA_API_KEY "nvapi-..." on Windows, or export NVIDIA_API_KEY=... on macOS/Linux, then restart Claude Code.
• Unknown model key → the server passes it through verbatim; surface the upstream error and suggest mcp__nvidia-models__list_models to find the right short key.
• Rate / quota errors → suggest waiting or routing to a cheaper model (e.g., llama-3.1-70b instead of llama-3.1-405b).

Bounds enforced server-side

The MCP server rejects requests that exceed: prompt 32,000 chars, system 8,000 chars, max_tokens 1–8192, temperature 0–2, models list 1–12 entries. Don't try to bypass — if the user asks for a longer prompt, chunk it first.