MCP Hub
Back to servers

ckg-mcp

Structured domain knowledge for AI agents. 42x more accurate than RAG, 11x fewer tokens.

Registry
Updated
Apr 27, 2026

Quick Install

uvx ckg-mcp

ckg-mcp

mcp-name: io.github.Yarmoluk/ckg-mcp

Compact Knowledge Graph MCP server. Pre-structured domain knowledge as a routing layer for agent stacks — 42× more efficient than RAG on structural queries.

License: MIT Python 3.10+ MCP Compatible

Built on the CKG Benchmark — 45 domains, 7,928 queries, fully reproducible results.


What It Does

Drop CKG into your agent stack as an MCP tool. Instead of retrieving text chunks and hoping the LLM infers structure, CKG gives agents pre-compiled dependency paths, prerequisite chains, and concept relationships — directly from a structured graph.

SystemMacro F1Tokens/queryHallucination Rate
CKG0.4712690%
RAG0.1232,982Variable
GraphRAG0.1203,450Variable

Install

pip install ckg-mcp

Claude Desktop Configuration

Add to your claude_desktop_config.json:

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

Tools

ToolDescription
list_domains()List all available CKG domains
query_ckg(domain, concept, depth)Extract subgraph — prerequisites + dependents
get_prerequisites(domain, concept)Full prerequisite chain to root
search_concepts(domain, query)Find concepts by name

Bundled Domains (v0.1.0)

DomainConcepts
calculus105
algebra-180
chemistry95
biology88
linear-algebra72
data-science-course91
economics-course78
glp1-obesity90

More domains available via Graphify.md — weekly-updated commercial CKGs for clinical, regulatory, legal, and financial domains.


Example

# In your agent — via MCP tool call
query_ckg(domain="calculus", concept="Taylor Series", depth=3)

# Returns:
## CKG: Taylor Series (calculus)

### Prerequisites (what you need to know first)
  - Power Series
    - Sequences and Series
      - Limits
  - Derivatives
  - Infinite Series

### Builds toward
  - Maclaurin Series
  - Error Estimation

Why Not RAG?

RAG retrieves text chunks and forces the LLM to infer structure. On multi-hop structural queries (prerequisites, dependency chains, category aggregation), that inference fails — F1 = 0.123 vs CKG's 0.471.

CKG is a pre-compiled routing layer: the dependency paths are already in the graph. BFS/DFS traversal, not similarity search. No hallucinations by construction.

Full benchmark: github.com/Yarmoluk/ckg-benchmark


License

MIT — Yarmoluk & McCreary, 2026. Commercial deployment → graphifymd.com

Reviews

No reviews yet

Sign in to write a review