dalao.ai

Graphify

An open-source tool that turns a codebase with its docs, configs and PDFs into a knowledge graph AI coding assistants query through a skill or MCP; Apache-2.0, with pre-relicensing MIT contributions still under MITOpen source

About Graphify

Graphify turns a codebase, with its docs, SQL schemas, configs and PDFs, into a queryable knowledge graph that AI coding assistants such as Claude Code, Cursor and Codex query through a skill or MCP instead of grepping files. Code is parsed locally with tree-sitter and no LLM; only docs and media use a model. As of 2026-10-26 the PyPI package graphifyy is at 0.9.95, released on GitHub 2026-10-24 01:59 (UTC+8). The engine is Apache-2.0, and portions contributed under MIT before the relicensing remain available under those terms; hosted Pro is $10 a month billed yearly, with a free tier.

EditorialUpdated 7 min read

What problem it solves

On a large codebase, AI coding assistants usually grep and read files one by one, which burns tokens and leaves the source of an answer unclear. Graphify first builds the codebase into a knowledge graph, and the assistant queries the graph instead of the files.

It works in two passes. Code goes through tree-sitter AST parsing, which is deterministic, uses no LLM and stays on the local machine. Docs, PDFs, images and video go to the model the assistant is running, or to a configured API key, for a semantic pass. The output is three files: graph.html for browsing, GRAPH_REPORT.md listing key concepts and surprising connections, and the full graph in graph.json.

Integration. The graphify install command detects the local AI coding assistants and registers a /graphify skill, and the graph can also be served over an MCP server for any MCP client.

The repository is Graphify-Labs/graphify, created 2026-04-03 23:49 (UTC+8), with v8 as its main branch. The PyPI package is graphifyy, with two y's, and needs Python 3.10 or later. Recent releases: v0.9.93 on 2026-10-21 00:29 (UTC+8), v0.9.94 on 2026-10-23 00:19 (UTC+8) and v0.9.95 on 2026-10-24 01:59 (UTC+8).

The publisher is Graphify Labs. The About page names Safi Shamsi as Founder and CEO; according to the YC company page, the company is in the Summer 2026 batch. Safi Shamsi is also the copyright holder in NOTICE and LICENSE-MIT.

The tool sends no usage pings, no crash reports and no analytics. Nothing to opt out of, because nothing is sent.

Graphify website

Core capabilities

Every edge carries its provenance. The README says each connection is tagged EXTRACTED, meaning explicit in the source, or INFERRED, meaning resolved by Graphify.

Query, path and explain. The graphify query command returns a scoped subgraph for a plain-language question, graphify path traces the route between two concepts, and explain describes a single concept, all against graph.json without rereading files.

Communities and hubs. Leiden splits the graph into subsystems with labels that need no LLM, and the most-connected concepts are listed. NOTE and WHY comments plus ADR and RFC citations become nodes linked to the code.

Coverage. The README says cross-file calls, imports, inherits and mixes_in relations are resolved across about 40 languages. Optional extras add PDF, Office documents, video transcription, SQL schemas, live PostgreSQL introspection, Terraform, and pushes to Neo4j or FalkorDB.

Hosted features. Graphify Cloud adds cross-repo graphs, PR review, formal verification, shared memory and Sentry and Jira connectors on top of the open-source engine, within the allowances on the pricing page.

Recent release. v0.9.95 merged 21 community pull requests, including a security fix that stages skill-install writes through randomly named temp files so a pre-planted symlink cannot redirect a write. The AST cache schema changed with it, so the first build after upgrading re-extracts.

Who uses it

The home page cites 135,072 GitHub stars, 9.3M PyPI downloads and 13.1M repositories indexed. The Y Combinator company page cites 116K+ stars and 6.5M+ downloads and lists several companies using it in production (as reported by the vendor). As of 2026-10-26 the GitHub API shows 135,099 stars and 12,896 forks; contributors number 342 by GitHub account, or 362 including anonymous ones.

README benchmarks: recall@10 of 0.497 on LOCOMO against 0.048 for mem0 and 0.149 for supermemory; 82.0% key-fact coverage on an ERPNext cross-tool task against a grep-and-read baseline of 70.8%; and zero LLM credits to build the graph (as reported by the vendor). Full tables and reproduction commands are in BENCHMARKS.md.

The About page lists coverage by Forbes, InfoQ and TechGig between September and October 2026.

How it compares

Against graph-based RAG: Microsoft GraphRAG is MIT-licensed and uses an LLM to extract a graph from unstructured text, so building the graph itself costs model calls. Graphify parses code deterministically through the AST, calls a model only for docs and media, and targets codebases.

Against language-server coding tools: Serena gives assistants symbol-level navigation through LSP; its licensing overview puts the Serena application under GPL-3.0-or-later and the SolidLSP component under MIT. Graphify builds a persistent graph ahead of time and serves it through a skill or MCP, with an Apache-2.0 engine.

Against long-document retrieval: PageIndex is a vectorless, reasoning-based RAG for long documents under MIT, while Graphify targets codebases and the docs around them.

The feature table on the Graphify site sets it beside CodeRabbit, Greptile and Sourcegraph, labels those three proprietary and labels Graphify open core.

Who it's for / Not for

A good fit for:

  • Developers using AI coding assistants on large codebases or monorepos who face heavy token use and ungrounded answers. Claude Code, Codex, Cursor, Gemini CLI, Aider, OpenCode and others have install flags.
  • Teams wary of code leaving their machines. A code-only graph needs no API key, graphify extract runs fully offline, and --code-only skips the docs and images that would need a model.
  • Teams that want shared multi-repo graphs, PR review and formal verification, offered in hosted Teams and Enterprise with optional self-hosting.

A poor fit for:

  • Setups that cannot let document content pass through third-party models. The semantic pass sends docs, PDFs and images to a model, and headless runs select a provider from whichever API key is already in the environment.
  • Environments limited to Python below 3.10.
  • Integrators who need stable interfaces. Versions are still 0.9.x, and the cache schema changes between versions.

Getting started

Installing. The README gives uv tool install graphifyy, or pipx install graphifyy, with pip also possible; the package name ends in two y's. A single run of graphify install then detects the local assistants and registers the skill; --project limits it to the current project and --platform targets a specific assistant.

Building a graph. Typing /graphify . in the assistant builds a graph for the current directory, and graphify extract . does the same from the command line, producing graph.html, GRAPH_REPORT.md and graph.json. Optional features come as extras such as graphifyy[pdf], graphifyy[mcp] and graphifyy[watch].

Official channels. A dedicated page states that graphify.net is not affiliated with Graphify Labs, and that the only official code sources are Graphify-Labs/graphify on GitHub and graphifyy on PyPI.

Pricing & license

The open-source engine is free under Apache-2.0 and needs no account. The hosted product has four plans:

  • Free: $0 for one developer, 3 repositories, 25,000 graph nodes per repo, 10 push builds per repo a day, 15 PR reviews per workspace a month and 1 formal verification run a day.
  • Pro: $10 a month billed yearly, or $120 a year; $15 billed monthly. Unlimited repositories and graph size, 400 PR reviews and 400 verification runs a day.
  • Teams: $20 per seat a month billed yearly for the first 100 teams, then $28; monthly billing is $29, then $40. At least 2 seats, with shared graphs and a dedicated Slack channel.
  • Enterprise: optional self-hosting, licensed per developer seat with custom pricing, SSO/SAML and GitHub Enterprise support.

Pro and Teams include a two-week trial that requires a card. Qualified non-commercial open-source projects under MIT or Apache licenses can use it free, and Pre-Seed and Seed startups with under $1M in revenue over 12 months get 50% off Teams for up to 6 months.

License: LICENSE is the Apache License 2.0. NOTICE states that portions contributed under MIT before the relicensing remain available under those terms, with the original text kept in LICENSE-MIT. The terms of service say open-source components follow their own licenses while the terms cover the hosted service. The JetBrains plugin, Graphify MCP for JetBrains, is proprietary.

Risks & limits

On local processing and data egress, the official sources say the following: the home page says the open-source engine keeps parsing on the machine, with no uploads and no telemetry; the README says nothing leaves the machine for code, while docs, PDFs and images are sent to the AI assistant for semantic extraction, and headless graphify extract selects a provider from whichever API key is set, in the order Gemini, Kimi, Claude, OpenAI, DeepSeek, Azure, Bedrock, Ollama.

Hosted data handling. The privacy policy says the hosted service sends code and content to two subprocessors, OpenAI and Anthropic, to produce results, and that their API terms rule out training on it. Ingested content is removed from active systems within 30 days of deletion or account closure, and backups expire on a rolling 90-day cycle. The terms of service say customer content is not used to train models. With optional cookies accepted, the website runs PostHog and Meta Pixel.

Lookalike site. An official page states that graphify.net is not affiliated with Graphify Labs, and that the only official code sources are Graphify-Labs/graphify on GitHub and graphifyy on PyPI.

Assistant count. The home page cites 17 supported assistants; the PyPI summary names 16, without Copilot; the README install table includes GitHub Copilot CLI.

Issue backlog. As of 2026-10-26 the repository has 716 open issues and 1,635 open issues and pull requests combined. Issue 3728, opened 2026-09-22 11:26 (UTC+8) and still open, reports that one failed build permanently locks a free hosted project with no way to delete the source, project or account.

Compliance. The site footer reads SOC 2 Type II, Engagement Started. According to the YC company page, the team size is 1.

Our verdict

What sets Graphify apart: deterministic tree-sitter parsing of code that stays on the local machine and calls no LLM; an EXTRACTED or INFERRED provenance tag on every edge; an Apache-2.0 engine that can run fully offline; and a graph that many AI coding assistants query through a skill or MCP.

Its limits: the semantic pass for docs and media goes through third-party models, and the hosted service sends code and content to OpenAI and Anthropic; versions are still 0.9.x, with the cache schema changing between versions; the YC company page lists a team size of 1; and the SOC 2 Type II status is Engagement Started.

Sources

Official

  1. Graphify - the code knowledge graph for AI coding assistantsgraphify.com · Accessed Oct 26, 2026
  2. Pricing - Free, Pro, Teams, or Enterprisegraphify.com · Accessed Oct 26, 2026
  3. About Graphifygraphify.com · Accessed Oct 26, 2026
  4. graphify.net vs graphify.com: which is official?graphify.com · Accessed Oct 26, 2026
  5. Graphify Terms of Servicegraphify.com · Accessed Oct 26, 2026
  6. Graphify Privacy Policygraphify.com · Accessed Oct 26, 2026
  7. Graphify Plugin Licensesgraphify.com · Accessed Oct 26, 2026
  8. Graphify Changeloggraphify.com · Accessed Oct 26, 2026

Updates & reading