A traversable knowledge graph of the NVIDIA AI stack served as an MCP tool. Your agent gets structured, source-verified answers in 269 tokens instead of 2,982.
CKG traverses declared edges. RAG embeds text, searches vectors, retrieves chunks. The graph already knows that TensorRT-LLM depends on CUDA and what that means — it returns the subgraph, not a wall of retrieved text. Benchmark: ckg-benchmark v0.6.2 · 28 queries · F1 0.471 vs RAG 0.123.
After 150 trial calls, your agent gets this response. The upgrade link includes your exact savings — not a generic pitch.
{
"error": "Trial limit reached (150 calls).",
"tokens_saved_in_trial": 418500,
"message": "You saved ~418,500 tokens during your
trial — roughly $1.26 in RAG API costs.
$10/year continues that.",
"upgrade": "https://ckg-nvidia-ai.onrender.com/upgrade?saved=418500"
}
The tokens_saved field accumulates across every call in your trial — real savings, not an estimate.
Works with Claude Desktop, Cursor, Windsurf, or any MCP-compatible agent. No API key needed for the free tier.
{
"mcpServers": {
"ckg-nvidia-ai": {
"url": "https://ckg-nvidia-ai.onrender.com/mcp"
}
}
}
Benchmark dataset: huggingface.co/datasets/danyarm/ckg-benchmark
The free tier is real — enough to build something. The trial shows you the savings. The annual key removes the math.
A GitHub star tells me it's worth building more CKGs. Feedback tells me what to build next. Both take 30 seconds.