Graphify

What it does

Pointing an agent at a large codebase or document pile means re-reading the same files for every question — which burns through tokens fast. Graphify scans the folder once, extracts concepts and relationships into a knowledge graph, and routes future queries through the graph. The repo claims 71.5× fewer tokens per query versus reading raw files on large corpora.

Usage examples

After install, the /graphify family of slash commands builds the graph, queries it, and pulls in external sources.

# 1. Build a graph — current folder or any specific path
/graphify .
/graphify ./raw                                       # run on a specific folder

# 2. Ask what connects two concepts
/graphify query "what connects attention to the optimizer?"

# 3. Trace the path between two nodes
/graphify path "DigestAuth" "Response"

# 4. Explain a concept in graph context
/graphify explain "SwinTransformer"

# 5. Add an external source (paper, webpage) to the graph
/graphify add https://arxiv.org/abs/1706.03762        # fetch a paper, save, update graph

The headline workflow is the Andrej Karpathy /raw folder pattern that the README cites — a single folder of papers, tweets, screenshots, and notes that graphify merges into one queryable graph. Once it’s built, “what connects attention to the optimizer?” or “trace DigestAuth to Response” run against the graph instead of re-reading raw files. That’s where the token savings come from.

Key features

• Multimodal input

  Code, SQL schemas, R/shell scripts, PDFs, markdown, screenshots, diagrams, images, and videos — all from a single folder. Claude’s vision capabilities pull concepts and relationships out of visual material too.

• App + DB + infrastructure in one graph

  Application code, database schemas, and infrastructure definitions land in the same graph. You can trace which functions read which tables and run on which containers in a single view.

• Persistent graph across sessions

  Results are saved to graph.json so you can query across sessions without rebuilding. SHA256 hashing reprocesses only the files that changed.

• Multiple export formats

  HTML interactive visualization, Obsidian vault, Wikipedia-style articles, GraphML, Neo4j Cypher, and SVG — formats that drop into other tools without conversion.

• Auto-sync

  Watch mode and git hook integration keep the graph current as code or docs evolve, no manual rebuild step required.

• Edge provenance

  Every edge is tagged EXTRACTED, INFERRED, or AMBIGUOUS, making the line between explicit fact and model inference visible for review.

• MCP server mode

  Pass --mcp to expose the graph as a tool to other agents or clients — pushes the skill from solo use into multi-agent setups.

Output files

Notes

• PyPI package is temporarily graphifyy — the proper name is being reclaimed, so the package has an extra y. The slash command stays /graphify.
• Python 3.10+ required — graphify install provisions the skill on your system.
• Edge provenance tags — areas with many INFERRED or AMBIGUOUS edges are worth reviewing manually before relying on them.
• MIT licensed.