🕸️ CKG — Compact Knowledge Graphs for Agent Stacks

Pre-structured domain knowledge your agents can traverse. Deterministic. Zero hallucination. 11× fewer tokens than RAG.

65×RDS over
RAG

3.4×higher F1
than RAG

11×fewer
tokens

7,928queries
tested

Benchmark & Paper MCP Server pip install ckg-mcp graphifymd.com

🧠

Orchestrator / Planner

Receives goal · plans task chain · queries CKG before dispatching workers

agent calls trace_upstream(concept) via MCP ↓

🕸️

CKG — Compact Knowledge Graph

Returns exact dependency subgraph · deterministic · every edge is a cited fact

you are here

worker agent carries its CKG subgraph as context ↓

⚙️

Worker Agents

Execute with ground-truth structural context · traceable · auditable

pip install ckg-mcp One command. Your orchestrator calls trace_upstream(concept) like any MCP tool. Nothing else changes. Zero friction

Try it live — query the graph

Context returned to the agent

Token cost comparison Click a tile above to see the live comparison

—

CKG tokens

vs

—

RAG equivalent

— savings

RAG vs CKG — what actually changes

With RAG

Agent receives goal

↓

Dispatches worker immediately

↓

Worker retrieves similar chunks via embedding search

↓

Worker guesses from probability — may hallucinate, misses dependencies

Probabilistic · Unauditable · Stale context possible

With CKG

Agent receives goal

↓

Calls trace_upstream(concept) via MCP

↓

CKG returns exact dependency subgraph · 11× fewer tokens

↓

Worker executes with ground-truth context · every hop cited · no guessing

Deterministic · Every edge citable · No hallucination

Add CKG to your stack — three steps

Step 1

Install the MCP server

pip install ckg-mcp

Python 3.10+. Runs locally. No cloud dependency. 52 domains included.

Step 2

Add to your MCP config

// claude_desktop_config.json { "mcpServers": { "ckg": { "command": "ckg-mcp" } } }

Works with Claude Desktop, LangGraph, AutoGen, any MCP-compatible orchestrator.

Step 3

Query before you dispatch

# agent queries CKG before acting context = trace_upstream( concept="Chain Rule", domain="calculus", depth=3 ) worker.execute(task, context=context)

Worker gets exact dependency subgraph. Deterministic. Every hop citable.

View on GitHub → Enterprise domains Free: Calculus, Data Science, GLP-1, Orchestration & more · Enterprise: Clinical, regulatory, legal, financial, codebase — graphifymd.com

Want a CKG for your domain?

The query interface is domain-agnostic — swap the CSV, everything else stays the same. We've built CKGs for clinical pathways, regulatory frameworks, financial instruments, legal doctrine, and software codebases. Works with any MCP-compatible orchestrator: LangGraph, AutoGen, Claude, or your own.

Get your domain →

Codebase Intelligence

Ran CKG on LangChain Core — 180 modules, 650 dependency edges. One traversal returns the blast radius of any module change. Coding agents know what's safe to touch before writing a line.

Life Sciences

GLP-1 clinical pathway — 125 nodes, 170 edges. Payer strategy agents get the full mechanism-to-outcome chain before generating a single HEOR argument. No hallucinated citations.

Any Domain

52 domains built to date. 30-day pilot: you define the domain, we build the graph, you benchmark it against your current RAG. If it doesn't beat your baseline, you don't pay.