Y Combinator's Garry Tan Spends Up to $200K Annually on AI Token Usage, Shares Open-Source Coding Agent Framework

Image: TechCrunch · Source

Garry Tan, CEO of Y Combinator, disclosed spending $100,000 to $200,000 per year on AI tokens to power personal and business AI agents. He openly shared gstack, his open-source AI coding agent toolkit that functions as a virtual engineering team.

Garry Tan, the President and CEO of the prominent startup accelerator Y Combinator, has revealed that he spends between $100,000 and $200,000 annually on artificial intelligence tokens. This significant expenditure supports his extensive use of AI agents, which assist him with a diverse array of tasks ranging from coding and research to managing personal affairs. In a forthcoming episode of Walter Isaacson's podcast The Builders, Tan explained how his AI agents function almost like a trusted companion, helping with day-to-day activities and anticipating needs proactively.

Tan elaborated that his AI agents have comprehensive access to his digital information, including emails, Dropbox files, notes, and even therapy records spanning two decades. These agents operate autonomously to help him navigate and manage various aspects of his personal and professional life. He described this arrangement metaphorically as a "Jiminy Cricket" constantly advising and aiding him.

In addition to personal use, Y Combinator employs an internal AI agent named Stray Light that manages company operations more efficiently by maintaining contextual knowledge of all founders backed by YC, communications, and applications received.

Despite his enthusiasm, Tan acknowledges that wide adoption of such personalized AI agents will depend on reducing high operational costs, noting that his own spending represents a "crazy privilege" not feasible for most people. He also expressed skepticism about AI agents managed by large tech companies, emphasizing the importance of user control and personalization in AI's future.

To contribute to the AI ecosystem, Tan has open-sourced his AI coding framework called gstack. Available on GitHub under an MIT license, gstack comprises a set of 23 Claude Code-based AI skills configured to simulate a virtual engineering team. Roles include CEO, engineering manager, designer, QA lead, security officer, and release engineer, among others. The framework is designed to support software development agile processes from planning to shipping code and integrates tightly to ensure continuity and quality.

Gstack aims to democratize access to sophisticated AI-driven coding assistance by providing founder and engineering teams with free, structured tools that automate rigorous reviews, testing, and deployment workflows. It runs on multiple agent platforms beyond Claude Code and supports macOS, Linux, and Windows, albeit with some platform-specific considerations.

Tan's sharing of his personal AI tooling and experiences with significant token spending illustrates both the transformative potential and current accessibility challenges of AI agent technology within software development and personal productivity realms. As costs decline and tools mature, such agent-driven workflows may become increasingly mainstream.