graphify

Anyone who has joined an unfamiliar codebase knows the problem graphify is built for: full-text search and grep show you where a string appears, but not how the pieces relate. graphify is an AI coding skill, invoked as /graphify across roughly 20 agents including Claude Code, Codex, Cursor, Gemini CLI and Aider, that turns a folder of code, SQL schemas, docs, PDFs and images into a single queryable knowledge graph. The output lives in a local graphify-out/ directory rather than a database server, so there is no Neo4j or running service to set up. Once built, the graph becomes a memory layer that the agent can traverse to answer relationship questions such as what connects authentication to the database, instead of repeatedly re-reading files.

graphify runs a layered pipeline. Code is parsed locally with tree-sitter grammars across many languages, extracting functions, classes, imports and call relationships without any API calls. Non-code files such as PDFs, Markdown, Office documents and images are sent to a configured model API for semantic enrichment, which identifies concepts and links the syntax misses, tagging each relationship as extracted, inferred or ambiguous. A community-detection step using the Leiden algorithm then clusters nodes and surfaces highly connected hub nodes and unexpected cross-module connections. What stands out is the breadth: 36 tree-sitter grammars, optional video and audio transcription via faster-whisper, and outputs ranging from an interactive HTML view to JSON, an Obsidian vault, Mermaid call-flow diagrams and GraphML or Cypher for external tools.

The natural audience is developers and teams who already work inside one of the supported agents and want a codebase or mixed document set to become navigable rather than searchable. Concrete uses follow from the design. New team members can pull a committed graphify-out/ directory and start from an architecture map instead of cold-reading files. Reviewers can run PR impact analysis to see which changes touch the same graph communities and order merges accordingly. Teams can auto-generate architecture diagrams as Mermaid call flows, or expose the graph over MCP so the agent queries it before falling back to file search. A headless extract command runs in CI and lets you pick a backend, including local Ollama, when data residency matters.

The honest caveats are real. The semantic step sends your non-code files to your agent model API, which adds cost and a privacy consideration even though code itself stays local via tree-sitter. The project is pre-1.0 at version 0.8.31, so formats, flags and APIs can still shift between releases, and the changelog already references behaviour changes like automatic deduplication of ghost duplicates. Python 3.10 or newer is required, and the README notes that a plain pip install can fail if the skill Python differs from yours, with uv recommended instead. The interactive HTML view also degrades past roughly 5000 nodes, pushing you toward JSON queries on large repositories. None of these is disqualifying, but they shape where it fits.

Against a managed graph database like Neo4j with its own ecosystem, graphify is a different proposition: it is not a database you operate but a local, agent-driven extraction and query layer that can still push to Neo4j or FalkorDB if you want one. For making an existing codebase or document set understandable, especially for onboarding, review and architecture overviews, it is a credible and unusually broad open-source option under the permissive MIT license. It is most attractive when you already use one of the supported agents and accept the model API cost for non-code files. Treat it with caution for production-critical pipelines given the pre-1.0 status, and pin a version if you depend on stable output formats.