What is gstack? - Garry Tan’s AI Engineering Team Explained
A plain-language guide to gstack — Garry Tan’s open-source toolkit that turns Claude Code (and other AI coding agents) into a structured virtual eng team with real browser QA.
You’ve heard people ship software with AI coding agents. gstack is one concrete setup for doing that with a process — not a blank chat box.
gstack is an open-source toolkit by Garry Tan (Y Combinator President & CEO).
It turns an AI coding agent — especially Claude Code — into a structured virtual product and engineering team.
Slash commands. Markdown playbooks. A real browser so the agent can see your app. Free, MIT license.
The project’s own pitch: “That is not a copilot. That is a team.”
01
The problem
Blank prompts vs process
02
How it works
Skills + sprint order
03
What you’d do
A simple taste-test path
04
How it fits
GBrain, agents, LLMs
1. What problem it solves
Talking to an AI coder with no process often looks like this:
- Vague asks
- Scope that keeps growing
- Skipped tests and docs
- “It works on my machine”
- The agent forgetting what you decided yesterday
gstack’s answer is a sprint process:
Think → Plan → Build → Review → Test → Ship → Reflect
Each step is a named slash command — a specialist with a checklist — not one freeform chat that tries to do everything.
Think
Office hours
Plan
CEO / eng / design
Build
You + the agent
Review
Bugs + design
Test
Real browser QA
Ship
PR + deploy
Reflect
Retro + learn
For PMs and founders
gstack is not a no-code app. It assumes you (or your agent) work in a git repo with an AI coding tool. The product idea is still useful if you never install it: process beats a blank prompt.
Who it’s for
Per the project README:
- Founders and CEOs who still want to ship
- First-time Claude Code users who want roles instead of a blank prompt
- Tech leads who want review, QA, and release habits on every change
2. How gstack works
Skills = specialist playbooks
A skill is a Markdown workflow you invoke like /office-hours or /qa.
Think of each one as a hat the agent puts on: CEO, designer, eng manager, QA, security, release engineer.
The root /gstack command is a router — it helps you pick the right skill.
The README markets a core set of about 23 specialists + 8 power tools. The repo keeps growing (docs, iOS, safety, GBrain helpers). Treat “23 + 8” as the marketed core, not a forever inventory.
A few skills worth knowing by name
| You run | Role (plain English) |
|---|---|
/office-hours | Product interrogation before you build |
/plan-ceo-review | Scope and strategy challenge |
/autoplan | Runs several planning reviews in order |
/review | Hunt for production bugs |
/qa | Test in a real browser, fix, re-check |
/cso | Security-minded audit |
/ship | Tests, push, open a PR |
/retro | Weekly reflection |
You can also talk in plain language (“test the website”, “security check”). Skills can activate from phrases, not only slash names.
Eyes in the browser
Coding agents often guess from code. gstack gives them a browser.
- Prefer Aside on macOS (your real sessions) when available — gstack detects it; it does not install Aside for you
- Otherwise it uses gstack’s own bundled Chromium
/browse and /qa are how the agent sees the product instead of only reading files.
Ethos in one line
From gstack’s own ethos: AI models recommend. Users decide.
It’s framed as an Iron Man suit — you stay the pilot.
Check your understanding
What is gstack mainly trying to add on top of an AI coding agent?
3. What a normal person would actually do
The README’s recommended taste test is short:
Install gstack into your coding agent
Follow the repo’s 30-second install (clone + ./setup). Primary path is Claude Code; setup also supports other hosts like Codex, Cursor, OpenCode, and more.
Run /office-hours
Describe what you’re building. Get a structured product challenge and a design doc — before code.
Run /plan-ceo-review
Pressure-test scope. Expand, hold, or cut before you waste a sprint.
Build, then /review and /qa
Implement. Ask for a staff-style review. Point /qa at a staging URL so the agent uses a real browser.
Stop and decide
You don’t need the full factory on day one. The taste test is enough to know if the process fits you.
Bigger loops (when you’re ready)
- Design:
/design-consultation→/design-shotgun→/design-html - Ship:
/ship→/land-and-deploy→/canary - Memory across days: optional GBrain via
/setup-gbrain
What gstack is not
Not a SaaS dashboard. Not autonomous employees. Not a guarantee that the agent is always right. Skills are opinionated workflows — they still need your judgment.
4. How it fits with GBrain and LLMs
GBrain (sister project)
gstack teaches agents how to code and ship with process.
GBrain gives agents durable memory you control.
They are separate repos and release cadences. gstack plugs into GBrain with /setup-gbrain and /sync-gbrain — it does not bundle GBrain inside itself.
Large language models
If you’re still fuzzy on what an LLM is, start here: How does AI work?.
gstack sits one layer up. The model still writes tokens. gstack is the outer playbook — roles, reviews, browser tools, ship steps — wrapped around that model inside a coding agent.
You
└─ Coding agent (Claude Code, Codex, Cursor, …)
├─ LLM (writes the next pieces of text/code)
├─ gstack skills (process: plan → review → QA → ship)
└─ optional GBrain (memory that survives sessions)Quick recap
- gstack = process for AI coding agents, not a blank chat.
- Skills are specialist roles you invoke with slash commands.
- Browser QA lets the agent see the live product.
- You stay in charge — recommend / decide.
- Pair with GBrain when you want memory that sticks.
The takeaway
gstack is Garry Tan’s open-source “virtual eng team” for AI coding agents: think, plan, build, review, test, ship, reflect — with real browser eyes. Process over vibes.
References
- gstack on GitHub — source, install, skill list
- gstack on DeepWiki — browsable code explainer
- GBrain on GitHub — sister memory project
Related on Traction Growth
- What is Jev? — a typed decision model for routing, scoring, and software gates
- What is GBrain? — memory you control for the agent you already use
- How does AI work? — plain-English LLM explainer
- Home — product and growth learning from real systems
- Open source product growth case studies — deeper OSS explainers
What is Jev? - TypeSafe AI’s Decision Model Explained
A plain-language guide to Jev — how it turns messy language into bounded judgments, how it fits with LLMs and code, and what remains unproven.
What is GBrain? - Personal AI Memory Explained Simply
A plain-language guide to GBrain — Garry Tan’s open-source knowledge brain that gives the AI agent you already use durable memory with sources, corrections, and cited answers.