An Andrej Karpathy folder of random papers, tweets, screenshots, and raw notes is a mess that every AI founder recognizes. Graphify just turned that mess into the most useful developer tool of July 2026. The open-source Python project hit 90,000 GitHub stars with 8,379 new stars in a single week by solving a problem that every AI coding assistant user hits within their first session: AI tools are blind to the structure of your codebase. They read everything, pay for every token, and still miss the connections between files, papers, and diagrams that a human would spot instantly.

Graphify builds a persistent, queryable knowledge graph from any folder of inputs. Code, SQL schemas, documentation, research papers, screenshots, whiteboard photos, even images in foreign languages. It uses Claude vision to extract concepts and relationships from visual content, Tree-sitter for AST-level code parsing, and NetworkX with Leiden community detection to surface the connections that matter. The result is a graph that reduces token consumption per query by 71.5x on a mixed corpus against reading raw files, and it re-indexes a 2,900-file codebase in under two seconds.

What Graphify Actually Does

Type /graphify . in Claude Code and the tool reads your entire current directory. It parses every code file through Tree-sitter for AST extraction, runs every PDF through citation mining and concept extraction, and passes every image through Claude vision to identify concepts and relationships. It then builds a graph using NetworkX and the Leiden clustering algorithm from graspologic, outputs an interactive HTML visualization, an Obsidian-ready vault, and a persistent graph.json that survives across sessions.

The graph tags every edge with one of three labels: EXTRACTED (found in the source), INFERRED (derived by the LLM), or AMBIGUOUS (uncertain). This honesty about provenance is a design choice that separates Graphify from the growing field of code intelligence tools. No black boxes. You always know whether the connection between your database schema and your API handler was read from the code or guessed by the model.

The output includes a GRAPH_REPORT.md with god nodes (the highest-degree concepts everything connects through), surprising connections ranked by composite score, and suggested questions the graph is uniquely positioned to answer. For the Andrej Karpathy corpus of 52 files including repositories, five papers, and four images, the token reduction hit 71.5x. On a smaller corpus of four files, the reduction was 5.4x. The savings scale with corpus size.

Install and First Use

Graphify requires Claude Code and Python 3.10 or higher. The install is a two-liner:

pip install graphifyy && graphify install

The PyPI package is temporarily named graphifyy while the graphify name is reclaimed, but the CLI command remains graphify. For macOS users with externally managed Python environments, pipx install graphifyy is recommended. After install, open Claude Code in any directory and run:

/graphify .

The tool supports several modes. The --deep flag runs more aggressive inferred edge extraction for larger codebases. The --update flag re-extracts only changed files and merges them into the existing graph, which is critical for active development workflows. The --watch flag runs a background process that auto-syncs the graph as files change, with instant AST-only rebuilds on code saves and a notification for LLM re-pass on doc or image changes.

There is also a git hook:

graphify hook install

This installs a post-commit hook that rebuilds the graph after every commit. No background process required. It triggers once per commit, works with any editor, and coexists with existing hooks.

How Graphify Compares to Alternatives

Code-review-graph, the 20K-star tool we covered last week, focuses on the narrow problem of reducing context during AI code reviews. It builds a Tree-sitter AST graph, stores it in SQLite, and communicates via MCP with AI coding tools. Graphify solves a broader problem: it ingests not just code but PDFs, images, videos, and documentation, and is designed as a multi-purpose knowledge base for agents rather than a review-specific tool.

Apache Ossie, the Apache incubator project for semantic metadata, standardizes how metadata is exchanged between platforms. It is a specification with converters for dbt, GoodData, and Salesforce. Graphify is a local-first tool that actually builds the graph instead of defining the schema format. The two are complementary. Ossie defines the language; Graphify speaks it.

Where Graphify wins is in breadth of input types and multi-agent compatibility. It works with Claude Code, Codex, OpenCode, Cursor, Gemini CLI, and any tool that can call a CLI or MCP endpoint. The --mcp flag starts an MCP stdio server, making Graphify usable as a backend for any agentic workflow. The --wiki flag generates a directory of markdown articles that any agent can navigate by reading files instead of parsing JSON.

Who This Is For

Graphify is for solo founders running complex codebases who want their AI assistant to understand the full picture without burning tokens on every file. It is for researchers who keep a raw folder of papers, notes, and screenshots and need an agent that can navigate it. It is for teams with monorepos where the connection between frontend components, backend APIs, database schemas, and deployment configs is invisible to any single AI tool reading files in isolation.

At 90,000 GitHub stars and accelerating, Graphify is the fastest-growing developer tool of July 2026 by a wide margin. The trend line suggests it is solving a need that every developer who uses AI coding tools has felt but could not articulate: the need for a persistent, queryable model of their codebase that does not reset every time they start a new conversation.