A local knowledge graph engine that transforms code repositories, documents, papers, audio/video into queryable structured graphs, designed as a skill for AI coding assistants.
Graphify is a local knowledge graph engine designed for AI coding assistants. It leverages tree-sitter for local AST parsing across 25+ programming languages, extracting functions, classes, and modules as graph nodes and call/import relationships as edges — all with zero API calls. It also supports semantic ingestion of non-code content including PDFs, Office documents, images, audio/video (via local faster-whisper transcription), and YouTube URLs.
The core pipeline follows detect → extract → build_graph → cluster → analyze → report → export, employing Leiden community detection for automatic clustering and hub node identification. Each inferred relationship is annotated with a confidence level: EXTRACTED, INFERRED, or AMBIGUOUS. Incremental update mode re-extracts only changed files, making it suitable for repositories with millions of files.
It integrates with Claude Code, Cursor, Gemini CLI and other AI coding assistants through platform-specific install commands, enabling assistants to automatically consume graph reports before responding. An MCP server exposes structured interfaces including query_graph, get_node, get_neighbors, and shortest_path. Export formats cover HTML visualization, GraphJSON, SVG, GraphML (Gephi/yEd), Neo4j Cypher, Obsidian vault, and Markdown Wiki.
Privacy is a core design principle: code parsing and audio/video transcription run entirely locally; only documents/PDFs/images are processed through the user's own AI API key. No telemetry, no tracking. Security mechanisms include URL whitelisting (http/https only), file redirect interception, content size limits, path validation, and tag sanitization. Git Hook integration enables automatic graph rebuilding on commit/checkout, and .graphifyignore supports file filtering. The PyPI package name is graphifyy (double y) while the CLI command is graphify, requiring Python ≥ 3.10.
Unconfirmed information: pyproject.toml shows v0.6.9 while the website lists PyPI v2.1.0 — version discrepancy cause unclear; no HuggingFace page found; no associated academic papers; enterprise features (million-file support, SSO, audit logs) are on a waitlist with details TBD.