By ByteBot
Your AI coding assistant has no memory of your codebase. Every session, it re-reads the same files from scratch — grepping through 50,000 lines of code like it has never seen your project before. This is not a model limitation. It is an architectural flaw in how AI assistants consume context. Graphify, an open-source tool from YC Summer 2026, is the fix — and 105,000 GitHub stars in five months suggests this problem was more painful than anyone officially admitted.
The Problem Graphify Is Solving
Claude Code, Cursor, and GitHub Copilot are stateless by design. Every query sends raw file contents to the model, which re-parses the same classes, imports, and schemas every single time. On a large codebase, a complex refactor query can burn 120,000 tokens in context loading before the model produces a single line of output. Multiply that across a team’s daily sessions and you are not paying for AI assistance — you are paying for your AI to re-read your code.
The existing fix — RAG (Retrieval-Augmented Generation) — retrieves semantically similar text chunks. For documentation search, that works. For code, it falls short. You do not want to know what looks like a function; you want to know what calls it, what it imports, where its schema lives. Semantic similarity does not answer structural questions. A knowledge graph does.
What Graphify Builds
Graphify converts your entire repository into a queryable knowledge graph using tree-sitter — a deterministic AST parser that runs fully locally, makes zero API calls, and sends no code anywhere. It supports 40+ programming languages and maps not just code but SQL schemas, configs, PDFs, and documentation into the same graph.
Every relationship is tagged as either EXTRACTED (explicitly declared in source — imports, function calls) or INFERRED (resolved by Graphify — implicit dependencies, structural patterns). That distinction matters: you can see exactly what is ground truth and what is a computed inference. Three output artifacts are generated:
- graph.html — interactive visual explorer of your codebase structure
- GRAPH_REPORT.md — human-readable architectural summary
- graph.json — machine-queryable file for AI assistants
When your AI assistant needs to answer a question about your codebase, it traverses the graph instead of loading raw files. The result is a fraction of the token spend for the same (or better) architectural understanding.
The Token Reduction Numbers — Honestly
Graphify’s headline claim is 71.5x token reduction. That figure comes from a specific benchmark — a 52-file corpus of structured repositories. Independent benchmarks on real Python codebases measured 7.3x. Other tested conditions land between 49x and 60x for complex daily tasks.
The realistic range is 7x to 49x for typical use, with 71x achievable on large, well-connected repositories. Even 7x matters — it means running seven AI sessions for the cost of one. If you are on Claude Code and already watching the September 14 limit reduction, token efficiency stops being a nice-to-have and becomes triage.
Knowledge Graph vs. RAG for Code
The key architectural difference: RAG retrieves by similarity. A knowledge graph retrieves by structure. For code navigation, these are not equivalent.
Ask an AI “which services call the payment processor?” — RAG returns chunks that mention payment processing. Graphify returns the explicit call graph. The knowledge graph answer is verifiable and deterministic. The RAG answer depends on what ended up in the vector index. For structural reasoning about large codebases, graph traversal wins on code; RAG wins on heterogeneous document retrieval.
Getting Started
Setup is three commands:
uv tool install graphifyy
graphify install
graphify .The first command installs the CLI. The second registers the skill with your AI assistant — Claude Code, Cursor, Codex, Gemini CLI, or any of the 20+ supported platforms. The third builds the knowledge graph for your current repository. After that, trigger it from inside your assistant with /graphify .
No account required. No API key. No code leaves your machine. Apache 2.0.
Worth It?
For codebases under 5,000 lines with a flat structure, the overhead of generating and maintaining a knowledge graph probably does not pay off. For anything larger — especially enterprise repositories where agentic coding is taking hold — Graphify starts looking like infrastructure, not a tool. The 107,000 GitHub stars reflect a genuine and widely-felt pain point that had no clean solution until now.
It is a YC S26 project, so expect rough edges and occasional schema drift after major refactors. But the core concept is sound, the implementation is local-first and privacy-respecting, and the community reception points to something with staying power. Start with your largest, most interconnected repository and work down from there.
