Garry Tan Puts His Token Bill at $100,000 to $200,000 a Year, and His gstack Repo Has 139,600 Stars

Y Combinator President and CEO Garry Tan said he spends "about maybe $100[,000] to $200,000 a year on tokens" running AI agents, TechCrunch reported on Nov. 2. The remark comes from an episode of Walter Isaacson's podcast The Builders, due to air Nov. 3.

TechCrunch's Julie Bort wrote that the article contains no independent check of the figure. The models, providers and usage patterns behind the bill were not specified, and the range is hedged with "maybe."

What Tan described

Tan said he uses agents for work and coding, and for personal tasks. He said his agent has read his email, his Dropbox, his Apple Notes and "therapy notes going back 20 years," and he called it a "Jiminy Cricket on your shoulder at all times."

He also called the spending "too expensive right now" and "a crazy privilege that not everyone can do." Y Combinator runs an internal agent named Stray Light with what he called a "company brain" covering every founder the firm has backed and every application, per the report.

On his own range, the arithmetic is $274 to $548 a day over 365 days, by The Terminal's calculation.

1. Low end of range100K USD
2. High end of range200K USD

Source: TechCrunch, Nov. 2, 2026, quoting Tan on The Builders podcast

What gstack is, per its README

Tan has published his setup as gstack, an MIT-licensed set of Claude Code skills, which are Markdown files that define slash commands for a coding agent. The README casts them as a "virtual engineering team" with roles such as designer, reviewer, QA lead and release engineer. It says there is no premium tier.

The repository page showed 139,600 stars, 20,700 forks, 1,002 commits on main, 61 open issues and 15 open pull requests on Nov. 5. The About line says "23 opinionated tools," matching the 23 skills TechCrunch cites. The README's own tables list more than 23 skills and extra "power tools," so the two counts disagree on the page.

Agent host	Support level in README
Claude Code	Full
Codex CLI	Experimental
Cursor	Experimental
Gemini CLI	Experimental

Stars measure interest, not use. The repository does not report how many people run it or what they spend.

Cost controls in the repo itself

The README includes a /runaway-stopper skill that targets agent threads over 500 million tokens. Another skill, /eval-plan, requires a pre-approved budget before paid evaluations run. Telemetry is off by default and the update check is on, according to the README.

The issue tracker shows the project policing itself. Open issue #3987 reports that the gstack-redact tool flags ordinary property access as hostname leaks. Issue #3905 concerns binary files passing a secret-line rule in review exclusion, and #3924 covers a filesystem boundary that makes Codex skip a repository's own agents/ folder. Those are titles only, and the issue bodies were not read for this report.

Tan's description of giving an agent mailbox-level access sits next to those open security-adjacent issues. The report does not address privacy or security of that access. Tan has argued for a personal agent "that you control, that you program," according to TechCrunch, and the episode's air date of Nov. 3 is the first point at which his full remarks can be checked. For related agent tooling, see claude-mem's hooks and cloud sync and the Hugging Face intrusion record Geoffrey Hinton called a warning shot.