Searching for the AI YC interview with Gstack agents usually means one thing: you saw Y Combinator president Garry Tan talk about turning Claude Code into a virtual engineering team, and you want to know if the tool behind it is real and worth installing. It is — Gstack is a free, open-source skill pack, and this guide covers what Tan actually said, what the repo actually contains, and the exact steps to set it up yourself.
Short answer: Gstack is Garry Tan's open-source Claude Code add-on — role-based slash commands (CEO, engineering manager, QA lead, security officer, release engineer, and more) built to stop agents from "wandering and guessing." It's free under the MIT license, needs Claude Code, Git, and Bun v1.0+, and installs with one command in under a minute.
I cloned the repo straight from GitHub rather than trust secondhand write-ups, since the numbers floating around move fast enough that some of them already contradict each other. When I checked the GitHub API directly today, the project sat at 124,696 stars and 18,678 forks — up from the roughly 70,000 stars BigGo Finance’s coverage clocked back in April 2026, when it quoted Tan on why agents need "roles, process, and review," not just a bigger model. Here's what's actually in the box, what it costs to run, and where it can trip you up.
What you'll need
Gstack itself costs nothing — it's MIT-licensed and lives entirely as Markdown slash-command files. What you need is the stuff around it: Claude Code installed and signed in, Git, and Bun v1.0 or newer, since gstack's browser-automation skill is a compiled binary that Bun builds during install. The browser skill only runs on macOS and Linux; Windows users need Node.js too and lose that one piece. None of this needs a credit card beyond whatever Claude plan you're already paying for — Free doesn't include Claude Code, but Pro at $20/month does, and that's enough to run gstack's skills. Verified against Claude's own pricing page and the gstack README on the same day I wrote this.
Step-by-step: setting up Gstack
1. Confirm Claude Code and Bun are actually installed
This is where I got stuck first. I ran the clone command before checking for Bun, and ./setup failed immediately with "Error: bun is required but not installed" — it doesn't ship a fallback. Run bun --version before you do anything else; if it's missing, the setup script prints its own install command with a checksum-verification step rather than silently downloading a binary for you.
2. Run the one-line install
Open Claude Code in a terminal and run:
“ git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/.claude/skills/gstack && cd ~/.claude/skills/gstack && ./setup “
This clones the repo into your Claude Code skills folder and runs the setup script, which builds the browser binary and symlinks every skill so Claude Code can see them as slash commands.
3. Let Claude Code register the skills
Once setup finishes, tell Claude Code to add a "gstack" section to your project's CLAUDE.md listing the available skills — the README's own install prompt has Claude do this for you automatically, so you don't have to type out every command by hand.
4. Start with /office-hours, not a feature request
Tan's install flow is deliberate: /office-hours runs first and asks forcing questions about what you're actually building before any code gets written. Skipping straight to /autoplan or /ship defeats the point of the roles.
5. Run the review pipeline on real work
Once you have a branch with changes, /plan-eng-review locks the architecture, /review hunts for production-breaking bugs a passing CI run would miss, /qa opens a real browser to test the affected pages, and /ship handles the sync-test-push-PR sequence. Run them in that order the first few times before trying to shortcut it.
Example prompts you can copy
These are the actual slash commands from the repo, not paraphrased versions — copy them as-is once gstack is installed:
/office-hours— then describe your feature idea in plain language; it pushes back before you build./plan-ceo-review— run on a spec to find the "10x version hiding in the request," in the repo's own phrasing./review— run on any branch with uncommitted or unpushed changes to catch bugs CI wouldn't flag./qa https://your-staging-url.com— opens a real Chromium session and tests the pages that changed./cso— runs an OWASP Top 10 and STRIDE-style security pass before you ship anything customer-facing.
Common mistakes to avoid
The mistake I made first: cloning the repo before confirming Bun was installed, which killed setup on the first try. Second, running the browser-based skills like /qa and /browse on Windows and expecting parity — they need macOS or Linux, and Windows only gets a partial install. Third, jumping to /ship on day one instead of running /office-hours and /plan-eng-review first; the roles are sequential by design, and skipping the early ones just means /review catches problems that planning would have caught for free. Fourth, quoting Tan's productivity numbers — the "~810x" 2013-vs-2026 comparison — without reading his own methodology notes in the repo; he's explicit that it's a normalized-for-AI-inflation figure across his own repos, not a general claim about what gstack does for any team. Fifth, treating gstack as a chatbot instead of a workflow — the roles only add value when you actually run the sequence on real branches, not when you fire one slash command and expect a finished product.
Gstack vs. plain Claude Code
| Stage | Claude Code alone | Claude Code + gstack |
|---|---|---|
| Planning | One open-ended prompt | /office-hours + /plan-ceo-review + /plan-eng-review challenge scope and lock architecture first |
| Code review | You ask "review this" and get general feedback | /review runs a staff-engineer-style audit aimed at bugs that pass CI |
| QA | Manual testing, or none | /qa drives a real browser against your staging URL and auto-fixes what it finds |
| Security | Not built in | /cso runs an OWASP Top 10 and STRIDE pass |
| Shipping | Manual git commands | /ship syncs, tests, and opens the PR in one command |
| Cost | Included in any Claude Code plan | Free, MIT license, on top of your existing plan |
Tools that make this easier
Gstack only pays off once you're already comfortable in Claude Code — if you haven't set that up yet, my how to use Claude AI guide covers signup and the Pro-vs-Max decision first. If you're not sure Claude Code is the right coding agent for you at all, best AI tool for code tests it head-to-head against Cursor and OpenAI's Codex, and ChatGPT alternatives for coding covers the wider field if you want options beyond Anthropic's stack. Another open-source add-on worth knowing about in the same space is Unabyss for Claude, which tackles a different part of the Claude Code workflow. And if you're still deciding which AI tools are worth your time before adding a framework like gstack on top, your starter kit for AI is the place to start.
My take
Gstack is a genuinely useful structure, not just a viral README — the role-based slash commands map to real gaps. Nobody reviews their own pull request as rigorously as a second engineer would, and almost nobody runs a security pass before every single ship. The catch is that it demands the same discipline it enforces: run office hours before you build, actually read the diff the review skill flags, and don't skip straight to shipping just because the command exists. Treat it as a checklist that happens to be automated, and it earns the stars sitting on the repo today. Treat it as a magic button instead, and you will run straight into the same "plausible-looking code that silently breaks" problem Tan built the whole framework to solve in the first place.
For a solo builder or a small team already living inside Claude Code, the honest verdict is that gstack costs nothing but a few minutes of setup and a working Bun installation, and the downside if it doesn't fit your workflow is just as small — delete the folder and you're back where you started.
Last updated: July 27, 2026 · By Vishal Swami, Founder & Lead AI Reviewer, AISagely
Frequently Asked Questions
AI YC interview with Gstack agents: is it free?
Yes. Gstack is MIT-licensed and free to fork, modify, and redistribute — there's no premium tier or waitlist. You do need an active Claude Code plan to run it, and Claude Code requires at least the $20/month Pro plan, per Claude's pricing page.
How long does it take to set up Gstack?
The install itself takes under a minute if Bun and Git are already on your machine — one clone command and a setup script. Budget more time your first session, since /office-hours is meant to slow you down before you build, not speed you up.
What is the easiest way to try Gstack without committing to the whole workflow?
Install it, then run just /office-hours and /review on an existing branch. Those two skills show you the most value fastest without requiring you to restructure how you already work.
Does Gstack work with tools other than Claude Code?
The skills are built for Claude Code specifically, but the repo also documents setup for OpenClaw, which spawns Claude Code sessions and can call gstack skills that way. It isn't a general framework for other coding agents.
What was Garry Tan's point about AI agents needing "roles," not just prompts?
In his own words, quoted in BigGo Finance’s coverage of his remarks: "the way to get agents to do real work is the same way humans have always done it — as a team, with roles, with process, with review." His argument is that model intelligence isn't the bottleneck anymore; unstructured prompting is, because a model left to guess produces "plausible looking code that silently breaks."