Model Context Protocol server exposing NVIDIA NIM media tools for image generation, OCR, and video understanding.
Three tools with complete input schemas and descriptions. Naming follows verb_noun pattern (nvidia_<action>_<resource>). Descriptions are detailed (150-250 chars) and explain WHEN to use each tool. However, parameter descriptions lack constraint details (e.g., width/height accept 64-2048 but no guidance on common values or performance implications). Output schemas are not documented, LLMs cannot predict response structure for chaining. Error handling is minimal (generic error messages, no recovery guidance). No tool annotations (readOnlyHint/destructiveHint) despite clear risk profiles (WRITE vs READ_ONLY).
Generate an image from a text prompt via NVIDIA NIM Flux 2 Klein 4B (black-forest-labs). Saves the image to disk and returns the file path. Default output dir: ~/.cache/nvidia-media-mcp/output/. Use this for production-quality text-to-image renders when ComfyUI workflows are not needed.
Extract text from an image via NVIDIA NIM Nemotron OCR v1. Accepts a local image file path. Returns aggregated text plus per-detection confidence + bounding boxes. Use this for screenshot-to-text, document scanning, sign reading, etc.
Describe video content via NVIDIA NIM omni-modal models. Default model: nvidia/nemotron-3-nano-omni-30b-a3b-reasoning (Nemotron 3 Nano Omni 30B, reasoning + omni-modal). Pass a local video file path (MP4/MOV/WebM/M4V/MKV); the file is base64-encoded and sent inline. Soft cap: 50 MB per video — trim or downscale longer clips. Use this for short-clip understanding (sign-reading in video, scene description, action summary, etc.).
Output schemas not documented. LLMs cannot predict response structure (e.g., does nvidia_flux_generate return {path, size, format} or {image_path, metadata}?). This breaks tool chaining and forces agents to guess field names.
Parameter descriptions lack actionable constraints. width/height accept 64-2048 but no guidance on common values, performance trade-offs, or recommended defaults. 'Absolute path to a local image file' does not specify supported formats (PNG/JPEG/etc) or max file size.
No tool annotations (readOnlyHint/destructiveHint/idempotentHint). nvidia_flux_generate is marked WRITE (destructive, non-idempotent) but lacks annotation. Agents cannot determine retry safety or side-effect scope without explicit hints.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 55 | 2026-07-28+ | v2 |
Error handling is generic. errResult() returns raw error messages with no recovery guidance. 'Flux NIM error HTTP 401' tells agent nothing, should be 'Authentication failed. Verify NVIDIA_API_KEY is set and valid.'
No input validation. Accepts width/height outside documented range (64-2048) without early rejection. LLMs will pass invalid values; tool should validate and return 'width must be 64-2048, got 5000' rather than failing downstream.