After Karpathy shared his personal knowledge-base workflow, the open-source community shipped a productized answer within 48 hours. Graphify is a zero-configuration knowledge graph tool that automatically parses all the code, PDFs, images, and Markdown in any folder into a navigable knowledge graph, while pressing token consumption down by 71.5x. It suits developers and researchers who need to manage large amounts of research material, code repositories, or notes.

What Is Graphify

Graphify is the "complete form" of Karpathy's knowledge-base methodology. The original approach required manually organizing the raw/ directory and manually triggering processing, and repeatedly reading the raw files generated heavy token consumption. Graphify automates the entire workflow and cuts token usage by 71.5x with two-stage processing.

It has picked up 2k+ stars on GitHub; the PyPI package name is graphifyy (note the extra y).

Core Features

All-Modal Automatic Parsing

Just drop a folder in—no manual sorting needed:

• Code files: parsed locally via tree-sitter AST, with no LLM calls and zero token consumption
• PDF / Markdown: automatically split into text and semantic units
• Screenshots / flowcharts / whiteboard photos: Claude Vision is called to extract concepts and relationships

71.5x Token Savings

A two-stage pipeline: stage one does local AST extraction on code, calling no LLM at all; stage two runs a single semantic extraction pass over documents and images only. Combined with SHA256 caching, repeated runs process only the files that changed.

In a test scenario with 52 mixed files (a code repository + 5 papers + 4 images), per-query token consumption dropped by 71.5x.

No Vector Database Required

Clustering is based on graph topology (the Leiden community detection algorithm), not embeddings, which removes the deployment and maintenance cost of a vector database.

Installation and Usage

Requirements: Python 3.10+

# Install
pip install graphifyy && graphify install

# Enter the target directory and generate the knowledge graph in one command
/graphify .

When it finishes running, a graph.html appears in the current directory; open it in a browser to see the interactive knowledge graph.

Platform-Specific Installation

# Lobster platform (OpenClaw)
graphify install --platform claw

# Codex users need to enable parallel mode first
# Add under [features] in ~/.codex/config.toml:
# multi_agent = true

Tip: OpenClaw's support for multi-agent parallelism is still incomplete, so extraction runs sequentially only and is somewhat slower.

Incremental Updates and Watch Mode

# Watch mode: automatically re-parses after code changes
graphify --watch

# Incremental update: processes only new or changed files
graphify --update

It also supports installing Git hooks to rebuild the graph automatically after a commit or a branch switch, with no background process needed.

Verifying the Results

After running /graphify ., check whether the following files were generated in the current directory:

• graph.html: the interactive knowledge graph, open it directly in a browser
• Analysis reports and persistable data files

Every relationship carries a type annotation (extracted from the text / inferred by the model / ambiguous relationship) and a confidence score, with transparent, checkable provenance.

FAQ

• Why is the PyPI package named graphifyy? The name graphify was already taken when the author registered it, so graphifyy is the stopgap and may be renamed later.
• Codex parallel mode won't start? Check that ~/.codex/config.toml has multi_agent = true set.
• Too many graph nodes to read? Use --update for incremental updates, or use the HTML interface's filtering to focus on a specific community.

Project address: https://github.com/safishamsi/graphify/blob/v3/README.zh-CN.md