Garry Tan Spends $200K a Year on AI Agent Tokens, Open-Sources gstack

Executive Overview

Y Combinator CEO Garry Tan disclosed on The Builders podcast with Walter Isaacson that he spends between $100,000 and $200,000 annually on AI agent tokens. In parallel, Tan published gstack on GitHub — a free, MIT-licensed collection of 23 Claude Code skills that turns a single coding agent into a full virtual engineering team. The release and the spending figure together expose the real production cost of advanced AI agents and give developers a free framework to replicate the workflow.

Deep-Dive Features & Architecture

gstack redefines how a single coding agent operates by splitting it into specialized roles:

• /office-hours reframes the product idea and writes a design document.
• /review catches bugs that pass CI but fail in production.
• /qa opens a real browser, reproduces bugs, and fixes them.
• /cso runs OWASP + STRIDE audits.
• /ship runs tests and review, then opens a pull request.

gstack supports Team mode, which enables silent, network-failure-safe automatic updates with a stable release channel twice weekly (Monday and Thursday) at most once every 72 hours.

Installation requirements: Claude Code, Git, Bun v1.4.2+ (the setup rejects any version older than 1.3.3 to prevent bundled tools from reading .env files), and Node.js on Windows. On macOS, the Aside browser (macOS 15+) is recommended; browser skills, /make-pdf, and /diagram use it first with real recorded user sessions. The /cso skill requires a native toolchain; otherwise it remains unavailable.

On personal use, Tan said his agent "read all my email, all my Dropbox, all my Apple Notes, even my therapy notes going back 20 years." He also revealed that Y Combinator uses an internal agent called Stray Light that holds a "company brain" encompassing the full context of every founder YC has backed, everyone who has emailed them, and every application submitted.

Benchmark & Competitive Performance

gstack does not publish independent benchmark numbers, but its competitive value lies in being free under the MIT license compared with commercial agent-management tools. While Tan bears $100,000–$200,000 annually on tokens, a typical gstack user can start by running individual skills such as /review and /qa at a much lower token cost. The most notable structural advantage is support for eight different coding agents via the --host option, freeing users from lock-in to a single agent. Claude Code leads in enforcing security restrictions via hooks, while those restrictions remain advisory in experimental agents.

Industry Impact & Enterprise Adoption

The disclosure reframes AI agent economics: token spend is now a material operating cost for power users and startups. gstack's MIT license and multi-agent support lower the barrier for teams to standardize on role-based agent workflows without licensing fees. Y Combinator's internal Stray Light agent signals that venture firms themselves are embedding full-context AI agents into decision-making. For developers in the MENA region, gstack is globally available at no cost, with actual expenses tied to the underlying agent's token pricing.

Conclusion

Garry Tan's $100,000–$200,000 annual token spend is a hard data point on the real cost of advanced AI agents in production. gstack, free under MIT on GitHub, packages 23 Claude Code skills into a virtual engineering team with enforced security in Claude Code and experimental support across eight agents. The combination of a transparent cost figure and an open-source framework gives developers a concrete blueprint — and a clear budget line — for agent-driven engineering.

Media Source: TechCrunch AI | Official Developer Announcement: Official Source | Fact Verification & Analysis: AI Tools Oasis