Programmatic data access layer for LLM agents — navigate context, don't memorize it.

CGC connects AI agents to your databases, files, and vector stores. Instead of stuffing everything into the context window, your agent explores data on-demand: discovering schemas, sampling rows, chunking documents, extracting knowledge graphs, and querying across sources.

• Connect to PostgreSQL, MySQL, SQLite, filesystems, Qdrant, Pinecone, pgvector, MongoDB
• Discover schemas and structure automatically
• Sample data to understand what's in each table or file
• Chunk large documents into LLM-friendly pieces
• Search across files and databases with pattern matching
• Extract knowledge graphs from text and structured data (GLiNER2 unified model + 17 industry packs)
• Store extracted triplets in Neo4j, PostgreSQL AGE, or KuzuDB
• Query graph sinks with Cypher
• MCP server for Claude Desktop / Claude Code / Cursor / Windsurf
• HTTP API for integration with any tool or platform
• CLI for quick tasks and scripting

• Python 3.10+ must be installed on your system

pip install context-graph-connector

With graph extraction (GLiNER2 — requires ML models):

pip install context-graph-connector[extraction]

With legacy v1 extraction (GliNER + GLiREL + spaCy):

pip install context-graph-connector[extraction-v1]

With everything:

pip install context-graph-connector[all]

# Discover what's in a database
cgc discover sqlite ./mydata.db

# Sample rows from a table
cgc sample sqlite ./mydata.db users --n 10

# Run a SQL query
cgc sql sqlite ./mydata.db "SELECT * FROM orders WHERE total > 100"

# Extract knowledge graph from text
cgc extract "Steve Jobs co-founded Apple with Steve Wozniak in 1976"

# Extract from a file and store in a graph database
cgc extract-file ./report.pdf --sink kuzudb://./my_graph

# Chunk a large PDF for processing
cgc chunk filesystem ./docs report.pdf --strategy tokens:2000

# Start the HTTP API server
cgc serve

# Start the MCP server (for Claude integration)
cgc mcp

Claude Code (VS Code / CLI):

claude mcp add cgc -s global -- python -m cgc.mcp.server

Claude Desktop / Cursor / Windsurf — add to your config file (locations):

{
  "mcpServers": {
    "cgc": {
      "command": "python",
      "args": ["-m", "cgc.mcp.server"]
    }
  }
}

See the MCP Reference for step-by-step setup instructions for each editor.

from cgc import Connector

connector = Connector()

# Add a data source
from cgc.adapters.sql import SqlAdapter
connector.add_source(SqlAdapter("mydb", "sqlite:///data.db"))

# Discover schema
schema = await connector.discover("mydb")

# Sample data
rows = await connector.sample("mydb", "users", 5)

# Extract triplets
triplets = connector.extract_triplets("Elon Musk founded SpaceX in 2002")

CGC has a minimal core with optional extras for specific integrations:

cgc/
├── connector.py          # Main interface — Connector class
├── core/                 # Types: Schema, Query, Chunk, Triplet, Graph
├── adapters/
│   ├── sql.py            # PostgreSQL, MySQL, SQLite
│   ├── filesystem.py     # Local files (PDF, DOCX, CSV, etc.)
│   ├── vector/           # Qdrant, Pinecone, pgvector, MongoDB
│   └── graph/            # Neo4j, PostgreSQL AGE, KuzuDB (sinks)
├── discovery/            # Schema inference, relationship detection
│   ├── extractor.py      # Triplet extraction orchestrator (v1/v2 pipeline)
│   ├── gliner2.py        # GLiNER2 unified NER + relation extraction (v2, default)
│   ├── gliner.py         # GliNER v1 NER integration (legacy)
│   ├── glirel.py         # GLiREL relation extraction (legacy)
│   ├── router.py         # Industry pack routing (E5 embeddings, v1 only)
│   ├── industry_packs.py # 17 domain-specific extraction configs
│   └── structured.py     # Hub-and-spoke structured data extraction
├── cli/                  # Typer CLI
├── api/                  # FastAPI HTTP server
├── mcp/                  # Model Context Protocol server
├── session/              # Session tracking
└── security/             # API key auth, rate limiting

Contributions are welcome but this project is maintained on a best-effort basis. PRs may not be reviewed immediately. See CONTRIBUTING.md for guidelines.

Apache 2.0 — see LICENSE for details.