graphify.md / ckg-agentforce
Benchmark Paper Get access — $29/mo
CKG-AgentForce · Curated Knowledge Graph over MCP · v3

Model for language. Graph for truth.

CKG-AgentForce is a Curated Knowledge Graph for AgentForce Vibes — every quota rule, burn pattern, and Salesforce-documented constraint, pre-structured with sourced citations. Plug it into any MCP-compatible client — claude.ai, Claude Code, Cursor — and your agent stops re-reading docs and starts traversing declared relationships.

Get Developer Access — $29/month →
Cancel anytime · license key by email · patent pending
11×
fewer tokens
269
tokens / query
59
sourced edges, 43 nodes
30s
setup
v3 · live query output

Query the graph. Get the sources.

Every relationship returns a numbered citation back to the Salesforce doc where it was declared — at build time, not at query time.

ckg-mcp · agentforce-vibes · claude.ai connector
▸ query_ckg('Lifetime Token Allocation', 'agentforce-vibes', 4)
Lifetime Token Allocation [QUOTA] [1]
├─ Pro Request Limit [QUOTA] [1] ← REQUIRES(0.95)
├─ Token Ceiling [QUOTA] [1] ← REQUIRES(0.95)
├─ Credit Burn Rate [QUOTA] [3] ← REQUIRES(0.90)
├─ Quota Exhaustion [QUOTA] [1] ← REQUIRES(0.90)
├─ Context Compaction [STRATEGY] [2] ← REQUIRES(0.95)
└─ MCP Server Connection [INTEGRATION] [4] ← REQUIRES(0.90)
SOURCES
[1] help.salesforce.com/…/generative_ai_usage_einstein_requests
[2] developer.salesforce.com/…/five-pro-tips-agentforce-vibes
[3] help.salesforce.com/…/generative_ai_usage_metering_logic
[4] developer.salesforce.com/…/agentforce-vibes-claude-mcp
269 tokens · declared edges · check any relationship at the source
Screen recording · Jul 14 2026 · claude.ai

Recorded live in claude.ai.

User asks: "What's burning my AgentForce Vibes tokens and how do I fix it?" — Claude traverses the graph and returns sourced burn drivers with the fix chain.

claude.ai MCP tool execution — query_ckg returning result
01 · MCP call — query_ckg executes; declared edges returned, not inferred
Claude's graph-grounded answer with burn chains and traversal
02 · Graph-grounded answer — burn drivers + chain: Credit Burn Rate → Org Token Usage → Quota Exhaustion
What you get

One domain. Full graph. Works in 30 seconds.

agentforce-vibes · agentforce-developer
AgentForce Vibes CKG

43 nodes, 59 sourced edges. Every quota relationship, every burn pattern, every Salesforce-documented constraint — declared before your agent starts.

included in Developer plan
query_ckg → [1][2][3] + Sources
v3 Perplexity-style citations

Every query returns numbered source citations back to the Salesforce doc where the relationship was found. Auditable by construction.

declared at build time
graphifymd.com/api/mcp
claude.ai connector

Paste one URL and your API key into claude.ai Settings → Integrations. No install, no config file, no restart. Works in a new conversation immediately.

fastest path
pip install ckg-mcp · uvx ckg-mcp
Any MCP-compatible client

Cursor, Claude Desktop, Claude Code CLI, Windsurf — standard MCP config works everywhere. Same graph, same citations, same 30-node traversal, whichever client you prefer.

standard MCP
Setup

30 seconds, two options.

A
claude.ai connector (fastest)

claude.ai → Settings → Integrations → Add MCP Server. Two fields only — Name: CKG — AgentForce, and the URL with your key. Hit Add, start a new conversation.

https://www.graphifymd.com/api/mcp?key=YOUR-LICENSE-KEY
B
Cursor, Claude Desktop, or Code CLI

Install the package and register the server with your key — standard MCP config, works in any MCP-compatible client. Restart once.

pip install ckg-mcp && claude mcp add ckg uvx ckg-mcp --env CKG_API_KEY=your-key
✓
Verify either way

Run the query below. You should see the quota traversal tree with source citations — 269 tokens.

query_ckg('Lifetime Token Allocation', 'agentforce-vibes', 4)
Proof · benchmark v0.6.2 · 7,928 queries · 45 domains

Don't take our word for it.

Every number traces to the same public dataset: huggingface.co/datasets/danyarm/ckg-benchmark

0.471
F1 · same LLM, same questions
CKG0.471
RAG0.123
GraphRAG0.120

The difference is declared edges vs similarity retrieval.

11×
token reduction

CKG averages 269 tokens per query. RAG averages 2,982 on the same questions. The graph traverses declared relationships — agents stop re-reading docs every session.

HF · danyarm/ckg-benchmark · v0.6.2
110
lifetime AgentForce requests

Developer Edition orgs get 110 requests on Claude Sonnet 4.5 — one-time, non-refreshing. Salesforce's own fix: use MCP. CKG is the knowledge graph built for exactly this.

developer.salesforce.com · Vibes MCP Guide

Get the knowledge layer. $29 a month.

License key by email. Works with claude.ai, Claude Code, Cursor, and any MCP-compatible client. Cancel anytime.

Get Developer Access — $29/month →
benchmark v0.6.2 · F1 0.471 · 269 tokens/query · patent pending
© 2026 Graphify.md · patent pending
All plans Benchmark Paper Support