ckg-mcp 0.7.4

Compressed Knowledge Graph MCP server — 85 domains, 4x the F1 of RAG at 11x fewer tokens (~42x RDS). Works with Claude, GPT-4o, Gemini, Llama, Mistral, and any MCP client.

ckg-mcp

Give your agent the structure, not the search.

ckg-mcp serves Compressed Knowledge Graphs to any LLM via MCP — pre-structured, typed dependency graphs your agent traverses instead of text chunks it guesses from.

3.8× the F1 of RAG · 11× fewer tokens · works with Claude, GPT-4o, Gemini, Llama, Mistral, and any MCP client

Measured on the open CKG Benchmark — 45 domains, 7,928 queries. Re-run it yourself.

---

Quickstart — 30 seconds

pip install ckg-mcp        # Python ≥ 3.10
# or zero-install:  uvx ckg-mcp

Pick your client:

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 · any MCP client

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

Or with uvx (no global install):

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

LangChain / LangGraph / smolagents (Python)

from langchain_mcp_adapters.client import MultiServerMCPClient

client = MultiServerMCPClient({
    "ckg": {"command": "ckg-mcp", "transport": "stdio"}
})
tools = await client.get_tools()

OpenAI SDK

import subprocess, json
proc = subprocess.Popen(["ckg-mcp"], stdin=subprocess.PIPE, stdout=subprocess.PIPE)
# Use any MCP-over-stdio adapter — e.g. mcp-client-python or openai-agents-mcp

Restart your client, then try:

list_domains()
get_prerequisites(domain="calculus", concept="Taylor Series")
query_ckg(domain="nvidia-gpu-inference", concept="FlashAttention-3", depth=3)

---

What you get

6 tools. No database. No embeddings. No API key.

Tool	What it does
list_domains()	List all available domains. Call this first.
query_ckg(domain, concept, depth=3)	Dependency subgraph around a concept — prerequisites + dependents up to N hops
get_prerequisites(domain, concept)	Full prerequisite chain — every concept to understand first, in order
search_concepts(domain, query)	Find concepts by name (partial match, case-insensitive)
list_agent_blueprints()	List pre-built agent configs for specific use cases
get_agent_blueprint(use_case)	Full blueprint: domains, constraints, workflow, prompt template, LangGraph hint

Example session

list_domains()
→ Available domains (65 free / 85 pro): algebra-1, agent-reliability, calculus,
  context-as-a-service, databricks-unity, glp1-obesity, hipaa-compliance,
  nvidia-gpu-inference, payer-formulary, snowflake-horizon, ...

get_prerequisites(domain="calculus", concept="Taylor Series")
→ Prerequisite chain for 'Taylor Series' in calculus (20 concepts):
  Function → Limit → Derivative → ... → Power Series → Taylor Series

query_ckg(domain="nvidia-gpu-inference", concept="KV Cache", depth=3)
→ ## CKG: KV Cache (nvidia-gpu-inference)
  ### Prerequisites
  - Memory Bandwidth
    - HBM3 Memory (L2)
  - Transformer Attention
    - Scaled Dot-Product Attention (L2)
  ### Builds toward
  - PagedAttention
  - Continuous Batching
  - Speculative Decoding

list_agent_blueprints()
→ Agent blueprints (2):
    gpu-inference-optimizer: Diagnoses GPU inference bottlenecks and recommends optimizations
    context-as-a-service-advisor: Designs CaaS pipelines replacing RAG with CKG-based retrieval

get_agent_blueprint("gpu-inference-optimizer")
→ Full spec: required domains, constraints, 6-step workflow, prompt template, LangGraph state machine

---

Free vs Pro

	Free	Pro — $99/mo
Domains	65	85
Healthcare	—	HIPAA, CPT, ICD-10, drug interactions, payer formulary, modeling healthcare data
Enterprise data	—	Databricks Unity, Snowflake Horizon, PostgreSQL, AWS, Azure Purview, GCP Dataplex, dbt, OpenLineage
AI infrastructure	—	NVIDIA GPU inference, context-as-a-service, agent reliability, AI governance, token cost crisis
Agent blueprints	2 included	2 included + priority access to new ones
License	MIT	Commercial

Upgrade: graphifymd.com/pro — key delivered to your inbox, activate with one env var:

export CKG_API_KEY=your-license-key
# restart your MCP client — all 85 domains now appear in list_domains()

---

Benchmark

Three architectures. Same questions. Open methodology.

	CKG (this tool)	RAG	GraphRAG
Macro-F1	0.471	0.123	0.120
Tokens / query	269	2,982	3,450
Cost / query	$0.0010	$0.0106	—
F1 @ 5 hops	0.772	0.170	—
Fabricated edges	0 — by construction	variable	variable

• 3.8× the F1 of RAG at 11× fewer tokens
• ~42× higher RDS (Retrieval Density Score) — the compound efficiency metric
• ~10× cheaper full query set ($7.81 vs $76.23)
• Scales with depth: CKG F1 rises from 0.37 → 0.77 at 5 hops; RAG is flat. Retrieval has no traversal mechanism.
• GraphRAG ≈ RAG (0.120 vs 0.123) — the word "graph" isn't the win. Pre-structured, compiled graphs are.

Also validated by independent academic benchmark: arXiv:2603.14045 (Zarrinkia, Thomo, Srinivasan — U. Victoria / Santa Clara) finds 73–84% of Graph-RAG errors are reasoning failures, not retrieval failures — the problem CKGs solve by construction.

Reproducibility

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

Full benchmark paper →

---

How it works

A Compressed Knowledge Graph is a plain-text .csv DAG — concepts, typed dependency edges, taxonomy tags — 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 runs deterministic BFS/DFS over declared edges and returns the exact subgraph the agent asked for. Because the server only returns edges that are declared in the data, it cannot invent a relationship. The LLM still writes the answer in prose — the knowledge it reasons over is exact.

No graph database. No vector store. No retrieval pipeline. No inference at query time.

1M RAG tokens = 335 queries   (burning context at $0.013/query)
1M CKG tokens = 3,717 queries (11× compression — the same budget, 11× the coverage)

---

Agent Blueprints (new in v0.6.0)

Blueprints are pre-built, domain-locked agent specifications: which CKG domains to load, workflow steps, constraints, a ready-to-use system prompt, example queries, and a LangGraph orchestration hint.

get_agent_blueprint("gpu-inference-optimizer")
→ Required domains: nvidia-gpu-inference, context-as-a-service
  Constraints: only traverse declared edges, cite concept IDs, flag gaps
  Workflow: 1. list_domains → 2. query_ckg bottleneck → 3. trace prereqs →
            4. identify optimization path → 5. recommend with citations
  Prompt template: [ready to paste]
  LangGraph hint: StateGraph with 4 nodes: diagnose, trace, optimize, report

---

Bundled domains (65 free)

AI tools — claude-anthropic, claude-skills, conversational-ai, cursor, deepseek, gemini-api, grok-xai, kimi-moonshot, midjourney, moss, openai-platform, qwen STEM / math — algebra-1, calculus, chemistry, circuits, computer-science, data-science-course, digital-electronics, ecology, fft-benchmarking, functions, genetics, geometry-course, intro-to-graph, intro-to-physics-course, linear-algebra, machine-learning-textbook, pre-calc, quantum-computing, signal-processing, statistics-course Life sciences — biology, bioinformatics, dementia, drug-interactions, glp1-muscle-loss, glp1-obesity Pedagogy / tools — automating-instructional-design, infographics, microsims, prompt-class, tracking-ai-course, vercel-ai-sdk, langchain-core Business / society — art-of-war, blockchain, digital-citizenship, economics-course, ethics-course, it-management-graph, laudato-si, learning-linux, personal-finance, reading-for-kindergarten, systems-thinking, theory-of-knowledge, unicorns, us-geography, asl-book

Pro only (23): Healthcare · Enterprise data · AI infrastructure — see full list →

---

Compatibility

Model-agnostic — the graph is plain text, readable by any LLM:

LLM	Agent framework	MCP client
Claude (all tiers)	LangChain / LangGraph	Claude Desktop
GPT-4o / GPT-4	AutoGen	Claude Code
Gemini 1.5 / 2.0	smolagents	Cursor
Llama 3 / 3.1	CrewAI	Cline
Mistral	Haystack	Windsurf
DeepSeek	OpenAI Agents SDK	Any MCP stdio client

Python ≥ 3.10 · single dependency (mcp) · stdio transport · zero configuration

---

License

MIT. Source learning graphs derive from the McCreary Intelligent Textbook Corpus.

Citation

@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   = {v0.6.2. https://github.com/Yarmoluk/ckg-benchmark}
}

---

Links: Benchmark · Paper · Pro domains · Graphify.md · arXiv:2603.14045