Server data from the Official MCP Registry
Pre-structured knowledge graphs for AI agents: 3.8x RAG's F1, 11x fewer tokens, 0 invented edges.
Pre-structured knowledge graphs for AI agents: 3.8x RAG's F1, 11x fewer tokens, 0 invented edges.
This is a well-structured MCP server with clear security posture. It reads bundled knowledge graph CSV files with no external API calls, no credential requirements, and minimal attack surface. Code quality is good with proper input validation and error handling. Permissions are appropriately scoped for its purpose as a local knowledge graph query tool. Supply chain analysis found 3 known vulnerabilities in dependencies (0 critical, 3 high severity). Package verification found 1 issue.
5 files analyzed · 8 issues found
Security scores are indicators to help you make informed decisions, not guarantees. Always review permissions before connecting any MCP server.
This plugin requests these system permissions. Most are normal for its category.
Add this to your MCP configuration file:
{
"mcpServers": {
"io-github-yarmoluk-ckg-mcp": {
"args": [
"ckg-mcp"
],
"command": "uvx"
}
}
}From the project's GitHub README.
mcp-name: io.github.Yarmoluk/ckg-mcp
Give your agent the structure, not the search. ckg-mcp serves Compressed Knowledge Graphs — pre-structured, typed dependency graphs — to any MCP client. Instead of retrieving text chunks and hoping the model infers the relationships, your agent traverses declared edges: prerequisites, dependency chains, and category membership, returned as a tight subgraph.
On the open CKG Benchmark, this approach scores 3.8× the F1 of RAG at 11× fewer tokens — and unlike RAG, it cannot fabricate a relationship that isn't in the graph.
pip install ckg-mcp # or: uvx ckg-mcp (zero-install)
Claude Desktop — add to ~/Library/Application Support/Claude/claude_desktop_config.json:
{
"mcpServers": {
"ckg": {
"command": "ckg-mcp"
}
}
}
Claude Code (CLI):
claude mcp add ckg -- ckg-mcp
Cursor / Cline / Windsurf — same block in your MCP settings:
{ "mcpServers": { "ckg": { "command": "ckg-mcp" } } }
Prefer no global install? Swap the command for uvx:
{ "mcpServers": { "ckg": { "command": "uvx", "args": ["ckg-mcp"] } } }
Restart the client. Ask your agent: "Use the ckg tools — list the domains, then trace the prerequisite chain for 'Taylor Series' in calculus."
Four tools over 65 bundled domains (no database, no embeddings, no API key):
| Tool | Signature | What it does |
|---|---|---|
list_domains | list_domains() | List all available CKG domains. Call this first. |
query_ckg | query_ckg(domain, concept, depth=3) | Extract the subgraph around a concept — related concepts up to depth hops. |
get_prerequisites | get_prerequisites(domain, concept) | The full prerequisite chain — everything to understand first. |
search_concepts | search_concepts(domain, query) | Find concepts in a domain by name. |
list_domains()
→ calculus, circuits, machine-learning-textbook, glp1-obesity, payer-formulary, google-dataplex, ... (65)
get_prerequisites(domain="calculus", concept="Taylor Series")
→ Taylor Series → Power Series → Higher-Order Derivative → Derivative Function
→ … → Limit → … → Function (20 concepts — every edge declared in the graph, not inferred from prose)
query_ckg(domain="circuits", concept="RC Discharging", depth=2)
→ subgraph: RC Discharging ← RC Circuit, Capacitor Energy Storage, Initial Conditions ...
Three architectures, same questions, measured on the open CKG Benchmark (45 domains, 7,928 queries, fully reproducible):
| CKG | RAG | GraphRAG | |
|---|---|---|---|
| Macro-F1 | 0.47 | 0.12 | 0.12 |
| Tokens / query | 269 | 2,982 | 3,450 |
| F1 @ 5 hops | 0.772 | 0.170 | — |
| Fabricated edges | 0 by construction | variable | variable |
Don't trust the numbers — re-run them:
git clone https://github.com/Yarmoluk/ckg-benchmark && cd ckg-benchmark pip install -r evaluation/requirements.txt python evaluation/ckg_harness.py --domain calculus python evaluation/analyze_results.py
A Compressed Knowledge Graph is a plain-text .csv/.md DAG — entities, typed dependency edges, and taxonomy — that the LLM reads directly:
ConceptID,ConceptLabel,Dependencies,TaxonomyID
1,Function,,FOUND
2,Domain and Range,1,FOUND
4,Composite Function,1|3,FOUND
ckg-mcp does deterministic graph traversal (BFS/DFS) over these declared edges and hands the agent the exact subgraph it asked for. Because the server returns declared edges, not generated text, it cannot invent a relationship that isn't in the graph — that's what "0 fabricated edges by construction" means. (The LLM still writes the final answer in prose; the knowledge it reasons over is exact.)
No graph database. No vector store. No retrieval pipeline. Drop the server in, or drop a single .md into a system prompt.
Reference graphs — laudato-si, art-of-war, token-cost-crisis, agent-reliability, ai-governance, hipaa-ai Data catalog / governance — google-dataplex, aws-data-catalog, azure-purview, databricks-unity, snowflake-horizon Life sciences / clinical — glp1-obesity, glp1-muscle-loss, drug-interactions, dementia, icd10-metabolic, cpt-em-coding, hipaa-compliance, payer-formulary, modeling-healthcare-data, bioinformatics, genetics, biology STEM / math — calculus, pre-calc, algebra-1, linear-algebra, geometry-course, statistics-course, functions, intro-to-physics-course, chemistry, ecology, signal-processing, fft-benchmarking Engineering / CS — circuits, digital-electronics, computer-science, quantum-computing, machine-learning-textbook, langchain-core, claude-skills, data-science-course, it-management-graph, intro-to-graph AI / data / pedagogy — conversational-ai, prompt-class, tracking-ai-course, automating-instructional-design, microsims, infographics, modeling-healthcare-data Business / society / other — economics-course, personal-finance, organizational-analytics, ethics-course, theory-of-knowledge, systems-thinking, digital-citizenship, learning-linux, reading-for-kindergarten, us-geography, asl-book, moss, unicorns, blockchain
Need a domain we don't ship? Build your own CKG from a CSV, or ask about managed enterprise domains → graphifymd.com.
Works with any MCP client — Claude Desktop, Claude Code, Cursor, Cline, Windsurf — and any agent framework that speaks MCP (LangGraph, AutoGen, etc.). Model-agnostic: the graph is plain text, so it works equally with Claude, GPT, Llama, or a local model. Python ≥ 3.10, stdio transport, single dependency (mcp).
The open package ships 65 domains under MIT. Managed enterprise domains (clinical, regulatory, financial), weekly delta updates, and pilot engagements are available through Graphify.md.
MIT. Source learning graphs derive from the McCreary Intelligent Textbook Corpus.
@misc{yarmoluk2026ckg,
title = {Benchmarking Knowledge Retrieval Architectures Across Educational
and Commercial Domains: RAG, GraphRAG, and Compressed (Compact) Knowledge Graphs},
author = {Yarmoluk, Daniel and McCreary, Dan},
year = {2026},
note = {Pre-print in preparation. v0.6.2.}
}
Links: CKG Benchmark · Paper · Graphify.md
Be the first to review this server!
by Modelcontextprotocol · Developer Tools
Web content fetching and conversion for efficient LLM usage
by Toleno · Developer Tools
Toleno Network MCP Server — Manage your Toleno mining account with Claude AI using natural language.
by mcp-marketplace · Developer Tools
Create, build, and publish Python MCP servers to PyPI — conversationally.