Graphify

Graphify is an open-source (MIT) Python CLI and AI assistant skill that turns any repository into a queryable knowledge graph, entirely on-device. Built by Graphify Labs (YC S26) and led by co-founder Safi Shamsi, it installs in one command and integrates with 17 AI coding assistants including Claude Code, Cursor, GitHub Copilot, Codex, and Gemini CLI. The project reports over 90,000 GitHub stars and 3.2 million PyPI downloads as of July 2026.

What It Is

Graphify is a code intelligence tool that replaces file-grepping and fuzzy RAG retrieval with a real, traversable knowledge graph. It parses 36 programming languages using bundled tree-sitter grammars — deterministically, with no LLM call and nothing leaving your machine — and produces three output files: an interactive graph.html, a human-readable GRAPH_REPORT.md, and a machine-readable graph.json. Every relationship in the graph is tagged EXTRACTED (found directly in the AST), INFERRED (resolved by the model), or AMBIGUOUS (evidence points multiple ways), so every answer your assistant gives traces to a real path you can audit.

How the Graph Differs from RAG

Traditional RAG pipelines chunk code, embed it, and retrieve fuzzy top-k results by cosine similarity — a process that severs the call graph and forces the model to guess at connections. Graphify builds a typed entity-relationship graph and traverses it hop by hop. The practical difference: instead of stuffing whole files into context, the assistant pulls back only the nodes and edges relevant to a question. The Graphify homepage attributes community-reported results of 71.5× and 79× fewer tokens versus grepping-based approaches, with the 79× figure coming from a MemVerge blog post covering a 496K-token codebase with zero vector database in the stack.

Core Workflow

Installation takes about five minutes and runs entirely on-device:

• Install the package: uv tool install graphifyy (the PyPI package uses a double-y)
• Register the skill: graphify install — auto-detects Claude Code, Cursor, Copilot, and others
• Build the graph: /graphify . inside your assistant
• Query: graphify query "who owns billing?", graphify path "UserService" "DatabasePool", graphify explain "RateLimiter"

The MCP server (python -m graphify.serve) exposes 10 tools — including query_graph, get_node, get_neighbors, and shortest_path — over stdio or HTTP, so a single shared process can serve the whole team.

Platform and Language Coverage

Graphify supports 36 languages parsed on-device via tree-sitter, including Python, TypeScript, JavaScript, Go, Rust, Java, C/C++, C#, Ruby, Kotlin, Swift, PHP, Scala, Dart, Zig, and infrastructure languages like Terraform (HCL). Beyond code, optional extras handle PDFs, Office documents, Google Workspace files, images, and video/audio transcription (via faster-whisper, locally). SQL schemas and live PostgreSQL introspection are also supported. The graph can be exported to Neo4j, FalkorDB, GraphML, SVG, Obsidian vaults, or a markdown wiki.

Privacy and Deployment Model

Code parsing runs entirely locally — no telemetry, no account, no API key required for code-only corpora. The graph is a file on disk. Non-code files (docs, PDFs, images) optionally use whichever model backend the user configures: Ollama for fully local inference, or cloud providers (Anthropic, OpenAI, Gemini, DeepSeek, AWS Bedrock, Azure OpenAI) under the user's own keys. The MCP server can be self-hosted over HTTP for team-wide access. An enterprise layer covering merge-gate verification, graph-aware code review, an engineering digest, and a Jira connector is in early access.

Update: v0.9.19

The latest release is v0.9.19, published July 18, 2026, with active daily releases (v0.9.16 through v0.9.19 shipped July 14–18). The project was created in April 2026 and has shipped rapidly since, with the GitHub repository showing 90,850 stars and 8,860 forks as of the latest data. The active development branch is v8. The project traces its origin to an Andrej Karpathy post about an LLM-readable "wiki" for codebases, and the community has documented the lineage publicly.