Skip to content
Agentic Context Research

Graphify

TL;DR

Section titled “TL;DR”
  • Python CLI skill (/graphify) for Claude Code, Codex, OpenCode, Cursor, and others; installs via pip install graphifyy.
  • Three-pass pipeline: deterministic Tree-sitter AST extraction → local audio/video transcription (faster-whisper) → parallel LLM subagents for docs, PDFs, images.
  • Merges all extracted nodes/edges into a NetworkX graph; applies Leiden community detection for semantic clustering.
  • Outputs graph.html (interactive), graph.json (persistent, queryable across sessions), and GRAPH_REPORT.md (god nodes, surprises, suggested questions).
  • Reports 71.5× fewer tokens per query vs. reading raw files on a mixed 52-file Karpathy-style corpus (as reported).
  • MCP server mode available via pip install graphifyy[mcp] (serve.py).
  • SHA256 incremental cache: re-runs only process changed files.

What’s novel / different

Section titled “What’s novel / different”

Most codebase-intelligence tools handle source code only and produce vector embeddings or AST snapshots. Graphify is explicitly multi-modal: it ingests code, Markdown, PDFs, images, diagrams, and video/audio in a single pipeline and connects them into one graph. The Leiden clustering step surfaces cross-modal “god nodes” — high-degree concepts that span architectural decisions in code and reasoning in prose. Persistence via graph.json means the graph survives session resets, unlike tools that rebuild on each query. The design rationale (“why”) rather than just structure (“what”) is a stated first-class goal.

Architecture overview

Section titled “Architecture overview”

Core design

Section titled “Core design”

Three isolated passes run sequentially:

  1. AST pass — Tree-sitter extracts classes, functions, imports, call graphs, docstrings from 23 languages. No LLM.
  2. Media pass — faster-whisper transcribes video/audio locally using a domain-aware prompt derived from corpus god nodes. Cached after first run.
  3. Semantic pass — Claude subagents run in parallel over docs, PDFs, images, and transcripts to extract concept nodes and design-rationale edges.

The three pass outputs are merged into a NetworkX graph. Leiden algorithm (graspologic, Python <3.13 only) assigns community IDs. Analysis identifies god nodes (highest degree) and surprising cross-community edges.

Interface / API

Section titled “Interface / API”
  • CLI: /graphify <folder> within an AI coding assistant; also graphify <folder> from terminal.
  • MCP server: serve.py exposes MCP protocol; installed via [mcp] extra.
  • Query commands: /graphify query <text>, /graphify path <nodeA> <nodeB>, /graphify explain <node>.
  • Outputs: graphify-out/graph.html, graphify-out/graph.json, graphify-out/GRAPH_REPORT.md, graphify-out/cache/.
  • Ignore file: .graphifyignore (.gitignore syntax) to exclude paths.
  • Neo4j export: optional via [neo4j] extra.

Dependencies

Section titled “Dependencies”

Core: networkx, tree-sitter>=0.23.0 + 23 language grammars. Optional extras: mcp, neo4j, pypdf+html2text, watchdog, matplotlib, graspologic (Leiden; Python <3.13), python-docx+openpyxl, faster-whisper+yt-dlp. LLM: uses the assistant’s configured API key; no bundled model.

Scope / limitations

Section titled “Scope / limitations”
  • Leiden clustering (graspologic) is Python <3.13 only; users on 3.13+ cannot use community detection without the extra.
  • Token reduction claim is on a single curated 52-file corpus; no independent benchmark.
  • Parallel LLM subagent step sends semantic descriptions (not raw source) to upstream model — cost scales with corpus size and is unquantified.
  • No information on graph quality degradation on very large repos (100k+ files).
  • graph.json format undocumented; no stated stability guarantee.

Deployment model

Section titled “Deployment model”
  • Runtime: Python 3.10+; local process.
  • Language: Python.
  • Storage: local filesystem (graphify-out/); optional Neo4j for graph persistence.
  • LLM dependency: requires an external model API key (Claude, Codex, etc.); LLM calls made during semantic extraction pass only.
  • Network: outbound only for LLM API calls and optional URL ingestion; no telemetry.
  • License: MIT.

Self-reported metrics

Section titled “Self-reported metrics”
  • 71.5× token reduction — Karpathy mixed corpus: 3 GPT framework repos + 5 attention papers + 4 diagrams (~52 files, ~92k words); average query cost ~1.7k tokens vs ~123k naive (as reported, graphify.net).
  • 3.7k+ GitHub stars (as reported, graphify.net, 2026-04-13).
  • 23 languages via Tree-sitter AST (as reported, PyPI README).

Open questions

Section titled “Open questions”
  • Benchmark methodology: single self-curated corpus; “naive” baseline (reading all raw files) is an extreme lower bound. No comparison against embedding-based retrieval or other graph tools.
  • Graph quality at scale: no data on repos >1k files; community detection quality with Leiden on very large graphs is untested.
  • LLM cost: semantic extraction pass makes parallel LLM calls proportional to corpus size; no cost estimate provided.
  • MCP surface: serve.py MCP integration is underdocumented; tool names, schemas, and transport not described on the website.
  • Stability: version 0.4.10 with rapid iteration (10 patch releases within ~1 week); API/output format stability unconfirmed.
  • graph.json schema: no published schema; downstream tooling (Neo4j export) may not survive format changes.