Graphify Open-Source Tool: Building a Persistent Knowledge Graph for the Entire Repository for Coding Agents to Query
The Graphify open-source project proposes first building a persistent knowledge graph for the entire code repository, then letting coding agents query the graph to modify code, rather than re-running grep and reading files on every task round. It performs local tree-sitter AST parsing on the code, labels edges as EXTRACTED or INFERRED, and outputs artifacts to graph.json, graph.html, and GRAPH_REPORT.md under graphify-out/, after which query, path, and explain can be used for querying. The author points out that whether marketing figures such as a stable 80% cost reduction can be achieved depends on repository size, whether the graph stays fresh, and whether the agent actually connects to graph queries; connection methods include registering a skill via graphify install, or mounting graph.json into Claude Code, Cursor, and Codex via graphify-mcp.
Why read this
AI score 63
The original article breaks down the approach of building a graph for the entire repository and then making it available for coding agents to query, and points out that marketing figures depend on repository size and whether the graph stays fresh.
Topics
Body · AI translation
Graphify: Build a knowledge graph for the whole repo in one pass, so Agents grep less
@graphify's approach: first build a persistent knowledge graph for the entire repo, then have the coding agent query the graph to modify code, instead of grepping from scratch and blindly reading files every task round; it saves tokens and reduces getting lost; whether it can truly reliably cut costs by 80% and other marketing numbers depends on repo size, whether the graph stays fresh, and whether the agent is actually wired into graph queries.
Open-source project: https://t.co/WppLaXgFx8
Three key takeaways 1. Persistent knowledge graph, not a vector store: it runs a local tree-sitter AST on the code (deterministic, edges labeled EXTRACTED / INFERRED); docs/PDFs/multimedia can then go through a semantic layer. Output lands in graphify-out/ (graph.json + clickable graph.html + GRAPH_REPORT.md), after which you can query / path / explain without rescanning the whole repo each time. 2. Less repeated grepping: the slogan is "map once, not grep forever"; when the agent asks "how does the auth flow work" or "how many hops from A to B", it traverses edges such as calls/imports/inheritance on the graph instead of searching strings en masse. Community detection, god nodes, etc. help surface subsystem hubs first. 3. Wire it into the coding agent: uv tool install graphifyy → graphify install registers the skill, and in the assistant use /graphify .; you can also use graphify-mcp to attach an existing graph.json to Claude Code / Cursor / Codex, etc., using tools like graph_stats and query_graph as a context source. Cloud also has "always-on memory / cross-repo" features, which should be viewed separately from the open-source CLI.