Answer-Box: Graphify, a new open-source CLI tool, converts entire codebases—including code, docs, PDFs, images, and videos—into a local, queryable knowledge graph without using embeddings or vector stores. It works with over 20 AI coding assistants including Claude Code, Cursor, and Codex, and can be installed in under 30 seconds via `uv tool install graphifyy`. The tool parses code with tree-sitter AST for deterministic, fully local analysis, while other file types use the assistant's model for semantic enrichment.
How Graphify works
Graphify replaces traditional file-grepping by building a real graph of concepts and their relationships. Each connection is tagged as `EXTRACTED` (explicit in the source code) or `INFERRED` (resolved by graphify), giving developers full transparency into how the graph was constructed. The tool does not rely on vector indexes or embeddings—it creates a traversable graph that users can query with natural language, trace paths between two concepts, or ask for explanations of a single node.
After running `/graphify .` inside an AI assistant, Graphify outputs three files: `graph.html` (an interactive browser-based visualization), `GRAPH_REPORT.md` (a summary of key concepts and surprising connections), and `graph.json` (the full graph data for repeated queries). The graph can be queried anytime without re-reading the underlying files.
Installation and platform support
Graphify is available as a Python package on PyPI under the name `graphifyy` (double-y). Users can install it via `uv tool install graphifyy`, `pipx install graphifyy`, or plain `pip install graphifyy`. After installation, running `graphify install` registers the skill with the user's AI assistant. For project-scoped installs, the `--project` flag places skill files in the current repository, for example under `.claude/skills/` or `.agents/skills/`.
The tool supports over 20 AI coding assistants, including Claude Code, Cursor, Codex, Gemini CLI, and GitHub Copilot. Some platforms require additional configuration: Codex users need `multi_agent = true` in their config for parallel extraction, and Codex uses `$graphify` instead of `/graphify`. The `--platform agents` flag targets the generic Agent-Skills framework for cross-assistant compatibility.
Querying and confidence tags
Once the graph is built, users can run commands like `graphify explain "APIRouter"` to see a node's source location, community, degree, and all its connections with confidence tags. The `graphify path` command traces the shortest path between any two concepts—for example, from "FastAPI" to "ModelField" in three hops. Every edge carries a confidence tag (`EXTRACTED` or `INFERRED`), so developers know whether a relationship was read directly from the source or derived by the tool.
Graphify also supports optional hooks and instruction files that automatically nudge the assistant toward graph queries instead of raw file reads. On hook platforms like Claude Code and Gemini CLI, a hook fires before search-style tool calls. On instruction-file platforms like Codex and Cursor, persistent files like `AGENTS.md` or `.cursor/rules/` provide the same guidance. The `GRAPH_REPORT.md` remains available for broad architecture reviews.
Benchmarks and reproducibility
Graphify comes with a benchmark suite where every system was run on the same harness with the same model and budgets. A judge blind-validated against a second judge achieved 90.6% agreement (Cohen's kappa 0.81). Full per-system tables, code-intelligence results, and reproduction commands are documented in `BENCHMARKS.md`.