Graphify

Overview

Graphify turns entire codebases — source code, documentation, SQL schemas, configs, and PDFs — into queryable knowledge graphs that coding agents can reason over. Instead of forcing agents to read raw files sequentially (slow, context-hungry, and lossy), Graphify pre-computes a structural map of the codebase: every function, class, module, import, and cross-file relationship is captured as a graph with typed, confidence-tagged edges. Built by Graphify Labs (Y Combinator S26 batch) and reaching 109K+ stars, the tool has become the standard approach to codebase understanding for agent-assisted development.

How It Works

Graphify runs a two-stage extraction pipeline:

1. Code extraction (deterministic, local) — Tree-sitter AST parsing runs entirely on your machine with zero API calls. Supports 37+ programming languages. Every function call, import, class hierarchy, and type reference becomes a node or edge in the graph.

2. Semantic extraction (optional, LLM-backed) — Documentation, PDFs, images, and media are processed through your chosen LLM backend (Claude, Gemini, OpenAI, Ollama, or others) to identify cross-file relationships and implicit connections that AST parsing alone cannot capture.

The combined graph is then clustered using Leiden community detection, which identifies natural subsystems in the codebase and generates human-readable labels for each cluster. The algorithm also ranks “god nodes” — the most highly-connected concepts that represent architectural bottlenecks or core abstractions.

Every relationship carries a confidence tag: EXTRACTED (explicitly present in source), INFERRED (resolved by Graphify’s analysis), or AMBIGUOUS (multiple possible interpretations). This lets agents and developers distinguish hard facts from educated guesses.

Use Cases

• Codebase onboarding — New developers or agents get a structural map of an unfamiliar project in minutes instead of hours
• Architecture analysis — Identify god nodes, circular dependencies, and unexpected coupling between subsystems
• Scoped agent queries — Agents query the graph (graphify query, graphify path, graphify explain) instead of reading entire files, drastically reducing context usage
• Code review context — Understand the blast radius of a change by tracing its graph connections
• Documentation generation — The GRAPH_REPORT.md output serves as an auto-generated architecture document highlighting surprising connections

Getting Started

# Install graphify
npm install -g graphify

# Index your codebase
graphify .

# Query the graph
graphify query "What modules depend on the auth service?"
graphify path "UserController" "DatabasePool"
graphify explain "PaymentService"

In agent harnesses (Claude Code, Cursor, etc.), invoke the skill directly:

/graphify .

The skill installs platform-specific hooks that nudge agents toward querying the graph rather than reading raw files.

Outputs

• graph.html — Interactive browser-based knowledge graph visualization with node selection, community filtering, and search
• GRAPH_REPORT.md — Architecture summary highlighting subsystems, god nodes, surprising connections, and suggested queries
• graph.json — Portable, queryable graph file usable by MCP servers or external tools

Alternatives

• Aider repo-map — Aider’s built-in repository mapping (lighter-weight, less detailed, no graph queries)
• CodeGraph — Microsoft’s code graph tool for .NET codebases
• Sourcegraph — Enterprise code search and intelligence platform (SaaS, not a local skill)