Graphify Labs, fresh from its Y Combinator cohort, has released an open-source engine that translates entire codebases into queryable knowledge graphs—and the tool has already found its way into production environments at companies including incident-management platform Rootly and fleet-tracking giant Geotab.
The London-based startup's approach tackles a specific friction point in AI-assisted coding: context window overload. When coding assistants try to understand a large repository, they often resort to scanning thousands of files indiscriminately, burning through token budgets and missing structural relationships. Graphify's engine instead parses code locally using tree-sitter grammars, building a graph of functions, classes, imports, and their connections that assistants can query directly. The company says all parsing happens on-device, meaning no code needs to leave the developer's machine.
The traction has been notable, if hard to fully verify at this stage. The project's GitHub repository has accumulated over 111,000 stars, and PyPI download counts suggest broad experimentation. What's perhaps more telling is the production adoption: Rootly built a plugin that extends the tool beyond code to incident data, while Geotab used it to index millions of words from conference materials into a searchable knowledge base.
How It Operates
Installation follows the familiar pattern of modern developer tools. Engineers add Graphify through package managers like uv or pipx, wire it into their coding assistant of choice, then run a single command to generate the graph. The output comes in three forms: an interactive HTML visualization for human browsing, a markdown summary highlighting key patterns, and a raw JSON file that AI assistants consume.
Under the hood, the engine relies on tree-sitter, a parsing library that builds abstract syntax trees for 36 programming languages. It identifies structural elements and their relationships locally, without calling external APIs or large language models during the graph-construction phase. The company positions this as both a cost advantage and a privacy feature for teams working on proprietary codebases.
Integration support spans a range of AI coding tools, including Claude Code, Cursor, GitHub Copilot, and Aider, plus several autonomous agent platforms. That broad compatibility likely contributed to the rapid uptake among developers testing new workflows.
The Adoption Question
Rootly's public documentation describes how its customers can now point the tool at their incident histories to build graphs from operational data—a use case that stretches beyond pure code analysis. Geotab's experiment with conference content suggests the underlying graph-building approach generalizes to any structured knowledge domain, though the company's marketing remains focused on software engineering.
The startup also lists engineers at Tweddle Group and Superagent as users in its company materials, though independent confirmation of those deployments wasn't readily available in public channels. This kind of opacity is typical for early-stage enterprise adoption, where companies often pilot tools quietly before announcing partnerships.
One milestone stands out: the company claims Graphify became the first in its Y Combinator batch to cross 100,000 GitHub stars, a threshold that typically signals strong developer enthusiasm. Jared Friedman, a Y Combinator partner working with the company, noted the achievement in online comments.
Performance and Real-World Results
Benchmark numbers tell part of the story. The company published results showing 49.7 percent recall at 10 on the LOCOMO dataset and 76 percent accuracy on LongMemEval-S, according to documentation in their repository. They claim the graph-building process uses roughly 11 times less memory than competing tools and eliminates LLM costs during ingestion, since it relies on local parsing rather than API calls.
An enterprise version, available through early access, reported improved code-review scores compared to published baselines, though the company acknowledged limitations including small test sets and methodological differences that make direct comparisons tricky.
Third-party reviews painted a more complicated picture. One developer praised the streamlined user experience and documented reductions in token consumption. But Alex Rusin, writing in mid-summer, found that Claude Code frequently ignored the graph when handling well-scoped tasks, defaulting to its standard file-search behavior. "Did you use Graphify during that session? The answer was no," he wrote before uninstalling.
That disconnect between technical capability and actual usage hints at a broader challenge for developer tools: getting AI assistants to consistently leverage new infrastructure requires more than just API hooks.
Crowded Territory
Graphify enters a landscape where several projects are exploring similar territory through different architectural choices. CodeGraph uses SQLite with full-text search and continuous file monitoring. GitNexus offers deeper program analysis, including impact and taint tracking, but carries a noncommercial license that complicates enterprise use. GrapeRoot emphasizes forced retrieval to prevent agents from bypassing the graph entirely, addressing the exact problem Rusin encountered.
Synaptic, a commercial alternative, published its own benchmark showing higher quality scores than Graphify across a set of open-source projects. Vendor-run comparisons require skepticism, of course, but they signal competitive pressure in a category that's still defining its standards.
The Team and Business Model
Safi Shamsi, founder and CEO, holds a master's in data science from the University of Birmingham. His thesis explored knowledge-graph-powered retrieval for language models, work the company says demonstrated meaningful improvements in accuracy and hallucination reduction—though academic benchmarks don't always translate directly to production systems.
"Migrating a large codebase with AI is like changing a train's engine while it is moving," Shamsi wrote in the company's launch materials, a metaphor that captures both the opportunity and the risk.
The team size remains tiny. Y Combinator materials list two people, though LinkedIn profiles suggest a third employee may have joined. No seed round beyond Y Combinator's standard investment has been announced publicly. The company lists headquarters in both London and San Francisco, which likely reflects the distributed nature of early-stage startups rather than a formal office presence in either city.
Enterprise Push and Open-Source Tension
Graphify is now recruiting design partners for an enterprise product that runs self-hosted inside customer infrastructure and adds formal verification of code changes. The company offers a free tier for individual developers and per-seat pricing for teams, though specific dollar amounts remain unpublished.
The open-source version carries an Apache 2.0 license, though the repository contains both Apache and MIT license files—an inconsistency that could create confusion for commercial adopters evaluating legal risk. The company hasn't publicly addressed the discrepancy.
Whether Graphify can convert GitHub stars into sustainable revenue remains the central question. Developer tools with strong open-source traction often struggle to find the enterprise upgrade path that justifies venture-scale outcomes. The production deployments at Rootly and Geotab suggest real utility, but transforming that into predictable contracts will require solving the usage problem Rusin identified: making sure AI assistants actually use the infrastructure when it matters.
