wtf is ckg-nvidia-ai?

ckg-nvidia-ai is an MCP server that gives an AI assistant a structured map of the entire NVIDIA AI developer stack, covering 20 domains and around 1,000 nodes such as CUDA, TensorRT, NIM, NeMo, Riva, and Jetson. MCP, short for Model Context Protocol, is the standard that lets tools like Claude or Cursor connect to outside data sources. Instead of the AI guessing or searching broadly for how these NVIDIA tools relate to each other, it can traverse a graph of declared, typed relationships, such as which tool requires which other tool to run. The project calls this graph a Compressed Knowledge Graph, or CKG. The core idea is that when an AI agent needs to know something like what TensorRT-LLM requires to run on a certain NVIDIA GPU architecture, it normally spends a large number of tokens re-figuring that out through search or inference. With this graph already declared, the same answer can be found by traversing far fewer tokens, since the relationship is already recorded rather than re-derived. The README states this is roughly eleven times fewer tokens than a typical retrieval-based approach, with accuracy backed by an internal benchmark. The server is strictly read-only: it never writes, changes, or executes anything, only returns graph traversal results. It can be used two ways. The hosted option requires no installation and works directly in Claude Desktop, claude.ai, or similar tools by pointing them at a provided web address. The local option installs the package with pip and runs it as an MCP server through a command line tool called uvx, which also works with Claude Code, Cursor, Windsurf, and Cline. Each fact in the graph is manually reviewed by a person and marked with a confidence level, either high confidence and cross-checked against NVIDIA's own documentation, unaudited but plausible, or explicitly flagged as uncertain, so a user knows how much to trust any given answer. The project is distributed as a Python package and is free to use.