ckg-mcp · Model Context Protocol Server

Give your agent the structure,
not the search.

Pre-structured, typed dependency graphs your agent traverses instead of text chunks it guesses from. Works with Claude, GPT-4o, Gemini, Llama, Mistral — and any tool that speaks MCP.

4×
F1 vs RAG
11×
Fewer tokens
97
Domains
0
Hallucinated edges
pip install ckg-mcp
◈ Explore a CKG
⚙ Install
▤ Benchmark
✦ Pro Tier

Live GPU → LLM inference knowledge graph — 85 nodes, 9 tiers. Click a node to explore. This is what your agent traverses instead of a text chunk. Open full-screen →

Source: ckg-mcp domain nvidia-gpu-inference · 85 nodes · 9 tiers · graphifymd.com

30 seconds to first query
pip install ckg-mcp        # Python ≥ 3.10
# or zero-install:
uvx ckg-mcp

Claude Desktop

{
  "mcpServers": {
    "ckg": { "command": "ckg-mcp" }
  }
}

Claude Code CLI

claude mcp add ckg -- ckg-mcp

Cursor / Windsurf / Cline

{
  "mcpServers": {
    "ckg": {
      "command": "uvx",
      "args": ["ckg-mcp"]
    }
  }
}

LangChain / smolagents

from langchain_mcp_adapters.client import (
    MultiServerMCPClient
)
client = MultiServerMCPClient({
    "ckg": {
        "command": "ckg-mcp",
        "transport": "stdio"
    }
})
tools = await client.get_tools()
Then try
list_domains()
# → Available domains (68 free / 97 total): algebra-1, calculus,
#   langchain-core, nvidia-gpu-inference, hipaa-compliance ...

query_ckg(domain="calculus", concept="Taylor Series", depth=3)
# → Prerequisites: Power Series → Derivative → Limit ...
# → Builds toward: Maclaurin Series, Fourier Series ...

get_prerequisites(domain="machine-learning-textbook", concept="Backpropagation")
# → Chain: Function → Partial Derivative → Chain Rule →
#          Forward Pass → Loss Function → Backpropagation
Works with every LLM
ClaudeGPT-4oGemini Llama 3MistralDeepSeek LangChainLangGraphsmolagents CrewAIAny MCP client
Three architectures. Same questions. Open methodology.

45 domains · 7,928 queries · clone and re-run it yourself

MetricCKG (this tool)RAGGraphRAG
Macro-F10.4710.1230.120
Tokens / query2692,9823,450
Cost / query$0.0010$0.0106—
F1 @ 5 hops0.7720.170—
Full run cost$7.81$76.23—
Fabricated edges0 — by constructionvariablevariable

Benchmark v0.6.2 · open & reproducible · read the paper →

Also independently validated: arXiv:2603.14045 (U. Victoria / Santa Clara) finds 73–84% of GraphRAG errors are reasoning failures — the exact problem CKGs solve by construction. The graph doesn't guess — it traverses.
Reproduce it
git clone https://github.com/Yarmoluk/ckg-benchmark
pip install -r evaluation/requirements.txt
python evaluation/ckg_harness.py --domain calculus
python evaluation/analyze_results.py

68 domains are free forever. Pro unlocks 29 production domains — healthcare, enterprise data, and AI infrastructure.

Free
$0
forever
  • 62 CKG domains
  • STEM, life sciences, AI tools
  • All 6 MCP tools
  • Agent blueprints (2 included)
  • Healthcare & clinical
  • Enterprise data stack
  • AI infrastructure
pip install ckg-mcp →
Pro
$99/mo
billed monthly · cancel anytime
  • All 97 CKG domains
  • Healthcare: HIPAA, CPT, ICD-10, drug interactions, payer formulary
  • Enterprise data: Databricks, Snowflake, PostgreSQL, AWS, Azure, GCP, dbt, OpenLineage
  • AI infra: NVIDIA GPU inference, context-as-a-service, agent reliability, AI governance
  • Key delivered instantly — one env var to activate
Upgrade at graphifymd.com/pro →
How it works: Pay at graphifymd.com/pro → Stripe redirects you to a page showing your license key → set export CKG_API_KEY=your-key → restart your MCP client → list_domains() shows all 97 domains. No reinstall, no version change.