Carton MCP

⭐ 1 stars • 🕑 Updated 2026-07-10

Marketplace • Docs

📦 Auto-published from the monorepo • CHANGELOG • sancovp/carton-mcp

A Zettelkasten-style knowledge management system that provides both a core library for concept management and an MCP server for agent-driven knowledge operations.

Overview

Carton creates a dual-storage concept management system where ideas are stored as both Neo4j graph nodes and GitHub markdown files. The system auto-discovers relationships from text descriptions, manages missing concepts, and provides sophisticated graph querying capabilities.

This package provides both a core library for direct concept management and an MCP server for agent consumption.

Core Library Features

🏗️ CartOnUtils - Core Business Logic

• Neo4j Graph Operations: Execute read-only Cypher queries on :Wiki namespace
• Concept Network Analysis: Get connected concepts with 1-3 hop relationship depth
• Missing Concept Management: Detect, track, and auto-create missing concepts
• Duplicate Detection: Find similar concepts using textual similarity analysis
• GitHub Integration: Commit concept files and missing_concepts.md to repository

⚙️ ConceptConfig - Configuration Management

• Dual Storage Setup: Configure both GitHub (PAT, repo) and Neo4j (URI, credentials)
• HEAVEN Data Integration: Uses HEAVEN_DATA_DIR for local concept storage
• Environment-based Config: Loads settings from environment variables

📝 Add Concept Tool - Content Creation

• Auto-linking: Discovers concept mentions in descriptions and creates relationships
• Relationship Inference: Creates bidirectional links and inverse relationships
• File Generation: Creates structured markdown files in wiki/concepts/ directory
• Neo4j Integration: Stores concepts as nodes with normalized properties

MCP Server Features

🛠️ 7 MCP Tools

• add_concept: Create concepts with auto-relationship discovery
• query_wiki_graph: Execute read-only Cypher queries on :Wiki namespace
• get_concept_network: Explore concept relationships with depth control (1-3 hops)
• list_missing_concepts: Show concepts referenced but not yet created
• calculate_missing_concepts: Scan all concepts and update missing_concepts.md
• create_missing_concepts: Bulk create missing concepts with AI descriptions
• deduplicate_concepts: Find similar concepts using similarity thresholds

📋 4 MCP Prompts

• add_user_thought: Capture user quotes verbatim with topic attribution
• update_known_concept: Update existing concepts while preserving relationships
• update_user_thought_train_emergently: Track how thoughts evolved into insights
• sync_after_update_known_concept: Create sync concepts for version control

Installation

[Installation instructions pending PyPI publication]

Architecture

Dual Storage Model:

• GitHub: Markdown files in wiki/concepts/ structure with auto-linking
• Neo4j: Graph database using :Wiki namespace with properties n (name), d (description), c (canonical), t (timestamp)

Auto-Discovery System:

• Scans concept descriptions for mentions of other concepts
• Creates relates_to relationships automatically
• Tracks missing concepts and generates creation templates
• Maintains bidirectional relationships and inverse inference

Dependencies

• HEAVEN Framework for Neo4j utilities and base tool classes
• Neo4j Python driver for graph database operations
• GitHub integration for version control and wiki generation
• MCP protocol for agent tool exposure

License

MIT License - see LICENSE file for details.