gstack is Garry Tan’s personal Claude Code setup, open-sourced. Tan runs Y Combinator, and the pitch is that this is literally how he ships software with an AI agent: 23 slash commands plus a handful of power tools that turn Claude Code into a simulated engineering team. The reason it matters is not any single command but the opinionated workflow they encode. gstack is less a toolbox to pick from and more a fixed sprint you follow.
The idea: a sprint, not a menu
The commands map to phases of a product sprint:
- Think:
/office-hoursreframes your idea with a set of forced questions and writes a design doc. - Plan:
/plan-ceo-review,/plan-eng-review, and/plan-design-revieweach take a role (founder, eng manager, designer) and pressure-test the plan from that angle. - Build: you code, then
/reviewdoes a staff-engineer pass that auto-fixes obvious issues and flags the rest. - Test:
/qaopens a real browser, tests the running app, and fixes bugs it finds. - Ship:
/shipsyncs main, runs tests, audits coverage, and opens a PR;/land-and-deploymerges and verifies production health.
The defining behavior is that the agent is built to push back. /office-hours will tell you that you are actually building something different from what you described, rather than executing your prompt literally. That adversarial framing is what separates gstack from a passive assistant.
Install
# Single machine, ~30 seconds (needs Claude Code, Git, Bun)
git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/.claude/skills/gstack \
&& cd ~/.claude/skills/gstack && ./setupFor teams, ./setup --team enables auto-update with no vendored files and no version drift. Beyond Claude Code, ./setup --host <name> installs into Codex, OpenCode, Cursor, Kiro, Hermes, and others, each into that host’s own skills directory. You then add the gstack guidance to your CLAUDE.md.
When it fits, and when it does not
It fits technical founders and engineering leads who already live in Claude Code and want a complete, opinionated process rather than a pile of options. It fits less well if you want to cherry-pick one or two tools: gstack is designed as a flow, and using /review without the planning phases leaves value on the table. It is also heavy. There are 23 commands plus power tools, an optional GBrain vector layer, and a Bun dependency, so the learning curve is real. If you want a lighter skills set, look elsewhere.
How it compares
| Project | Shape | Note |
|---|---|---|
| garrytan/gstack | Opinionated sprint workflow, one author’s setup | Process-first, founder lens |
| obra/superpowers | Skills framework + methodology | More general, cross-agent |
| anthropics/skills | Official agent skills | First-party building blocks |
| farion1231/cc-switch | Manage multiple coding agents | Orchestration, not a workflow |
superpowers is the closest peer in spirit (an opinionated methodology) but is more general; gstack is unapologetically one person’s exact process. cc-switch solves a different problem (switching between agents). anthropics/skills is the neutral baseline you might assemble your own flow from.
Gotchas from the issue tracker
gstack ships multiple commits a day, and the open issues are a good map of where adoption hurts:
- Windows is rough.
gstack-learnings-logcould silently fail under git-bash, dropping all AI-logged learnings (#1950), and a watchdog flashed a console window every tick (#1952). - autoplan needs guardrails. Long in-place plan edits ran without git stash or branch isolation (#1959), and scope detection could silently skip a phase with no diagnostic (#1957).
- GBrain config is fragile. The local-versus-managed database detection has produced false “broken-config” states (#1964, #1917), so the vector layer needs hand-holding.
- Cross-agent upgrades drift.
gstack-upgradere-registered Claude Code but let Codex and OpenCode installs silently fall out of sync (#1925).
The pattern: the Claude Code experience on macOS or Linux is the polished path; Windows and the multi-agent and GBrain extras are where you will spend setup time.
A note on the claims
gstack leans on productivity numbers from its author (large multiples of lines shipped versus years past, services launched in weeks). Treat those as one developer’s self-report, interesting as motivation but not a benchmark. The durable value is the workflow itself, which you can read, fork, and adapt regardless of whether the headline numbers generalize.
FAQ
Is gstack free? Yes. gstack is MIT-licensed and open-source, with no premium tier or waitlist; it is one developer’s Claude Code setup published in full.
Does gstack work on Windows? It runs, but the macOS and Linux Claude Code path is the polished one. Several open issues (the learnings-log failure #1950, the console-window flash #1952) are Windows-specific, so expect rough edges there.
Do I need Claude Code to use gstack? Primarily yes, but ./setup --host <name> also installs gstack into Codex, OpenCode, Cursor, Kiro, and Hermes, each into that host’s own skills directory.
Can I use just one gstack command like /review? You can, but you leave value on the table. gstack is built as a connected sprint where the planning phases feed the review and ship steps, so it pays off most used as a flow, not a menu.
Related reading
For a more general skills methodology see obra/superpowers; for the first-party building blocks see anthropics/skills; and if your real problem is juggling several agents, farion1231/cc-switch.