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.
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.
| System | Macro F1 | Tokens/query | Hallucination Rate |
|---|---|---|---|
| CKG | 0.471 | 269 | 0% |
| RAG | 0.123 | 2,982 | Variable |
| GraphRAG | 0.120 | 3,450 | Variable |
Install
pip install ckg-mcp
Claude Desktop Configuration
Add to your claude_desktop_config.json:
{
"mcpServers": {
"ckg": {
"command": "ckg-mcp"
}
}
}
Tools
| Tool | Description |
|---|---|
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)
| Domain | Concepts |
|---|---|
| calculus | 105 |
| algebra-1 | 80 |
| chemistry | 95 |
| biology | 88 |
| linear-algebra | 72 |
| data-science-course | 91 |
| economics-course | 78 |
| glp1-obesity | 90 |
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