gstack - An AI programming workflow tool open-sourced by the CEO of YC - AiBoss

What is gstack?

gstack is an AI-powered programming workflow for Claude Code, open-sourced by YC CEO Garry Tan, that transforms AI assistants into virtual engineering teams. The tool includes 15 expert roles (such as CEO reviewer, staff engineer, QA manager, etc.) and 6 enhancement tools, all invoked via slash commands. gstack supports running 10-15 sprints in parallel and features real-world browser automation, dual AI cross-review, and automatic document updates.

The main functions of gstack

Product Restructuring:/office-hours Redefine the product by asking six mandatory questions, challenging assumptions and generating implementation solutions.
Planning Review:/plan-ceo-review Examining needs from the CEO's perspective/plan-eng-review Locking in the architecture and test matrix,/plan-design-review Rate content from 0 to 10 and detect low-quality content using AI.
Design System:/design-consultation Build a complete design system from scratch, generating realistic prototype diagrams and DESIGN.md documents.
Code review:/review Identify and automatically repair potential hazards in the production environment./investigate Systematic root cause analysis, forced to stop after three failures.
Design Repair:/design-review After auditing, I personally fixed the problem, submitted it atomically, and generated before-and-after comparison screenshots.
Quality Assurance:/qa Real browser testing, automatic repair and regression test generation./qa-only Only generate a report, do not modify the code.
Security Audit:/cso Perform OWASP Top 10 and STRIDE threat modeling to provide specific attack scenarios.
Release and Deployment:/ship One-click sync, testing, and PR opening;/land-and-deploy Deploy and verify production health after merging;/canary Monitor errors and performance after deployment.
Browser automation:/browse Controlling a real Chromium instance to perform a click screenshot./setup-browser-cookies Import real browser cookies.
Dual AI review:/codex Independent review by OpenAI Codex was introduced, and blind spots were discovered through cross-validation with Claude.
Team debriefing:/retro Generate weekly engineering reports that include personal data, release continuity, and testing trends.
Safety protection:/careful Warning before destructive commands,/freeze Lock the editing area./guard Combining the two provides complete protection.
productivity tools:/autoplan One-click connection between CEO, design, and engineering reviews;/document-release Automatically synchronize document and code changes;/gstack-upgrade It updates itself and displays the changes.

Key information and usage requirements of gstack

authorGarry Tan (President and CEO of Y Combinator)
position: An open-source toolset for transforming Claude Code into virtual engineering teams
Core data600,000 lines of code in 60 days, producing 10,000-20,000 lines per day, running 10-15 sprints simultaneously.
composition15 expert roles + 6 enhancement tools, all with slash command invocation.
Required tools：Claude Code, Git, Bun v1.0+
Windows AdditionalNode.js (Bun has Playwright compatibility issues on Windows, so it automatically falls back to Node.js)
Installation timeApproximately 30 seconds
ConfigurationThe CLAUDE.md file needs to include a gstack skills list.
Browser FunctionsOptional: Chrome/Arc/Brave/Edge for cookie import

gstack's core advantages

Structured processSupports transforming scattered AI prompts into a complete Sprint process (thinking → planning → building → reviewing → testing → releasing), avoiding chaotic output.
Role specializationThe system features 15 expert roles, each with its own specific responsibilities: the CEO reviews the direction, staff engineers find bugs, and QA uses a real browser to test the process, simulating real team collaboration.
Parallel capabilityIt supports running 10-15 independent sprints simultaneously, managed in the same way as the CEO managing the team, with important decisions taking precedence and the rest running automatically.
Real Browser:/browse Control the real Chromium instead of a simulator, and actually click, take screenshots, and process CAPTCHAs, so that the AI can truly "see" the interface.
Dual AI cross-validation:/codex We introduced OpenAI Codex for independent review, compared it with Claude to identify blind spots, and improved code quality.
Automatic repair closed loop:/qa Once a bug is discovered, it is automatically fixed, regression tests are generated, and re-verification is performed, completing the entire repair process without human intervention.

How to use gstack

Installation EnvironmentOpen Claude Code, paste the installation command, and Claude will automatically clone the repository, compile the binary file, and register all skills locally.
Configuration ProjectFor team sharing, copy gstack to the project. .claude/skills/ Directory and rerun setup, in CLAUDE.md Add a skill list for Claude to recognize.
Initialize the product (/office-hours)When you describe the product you want to create, AI will question your expression framework, challenge your underlying assumptions, extract hidden requirements, and generate design documents that automatically flow into downstream skills.
Planning review (e.g., plan-ceo-review)The process involves three layers of checks: CEO review to determine the direction, engineering review to lock in the architecture, and design review to score the design. Only after these three layers of checks are the solutions deemed feasible can development begin.
Code implementation (automatic or manual)After approving the plan, exit the planning mode. The AI will automatically write code based on the design document, or you can manually develop the code and then proceed to the review stage.
Quality review (/review + /qa):run /review Identify and automatically repair potential production hazards before resuming operation. /qa Open a real browser, click "Test," find bugs, automatically fix them, and generate regression tests.
Deployment (/ship):implement /ship One-click synchronization of the main branch, running tests, auditing coverage, pushing code, and automatically starting pull requests completes the release process.
Ongoing maintenance (/retro + /gstack-upgrade)Regular operation /retro Review team data and run /gstack-upgrade Update yourself to the latest version to get new features.

gstack project address

GitHub repositoryhttps://github.com/garrytan/gstack

Comparison of similar products to gstack

Dimension	gstack	OpenAI Codex	Devin (Cognition)
position	Claude Code's Virtual Engineering Team Workflow	OpenAI Official CLI Code Assistant	Fully Autonomous AI Software Engineer
Core Form	15 expert roles + 6 tools, slash command invocation	Single-session command-line tool	Independent cloud agent, end-to-end development
Workflow	Structured Sprint (Think → Plan → Build → Review → Test → Release)	Free dialogue, no fixed procedure	Autonomous planning and execution with minimal human intervention
Parallel capability	Supports 10-15 Sprints running simultaneously.	single session	Single task
Code review	`/review` + `/codex` Dual AI cross-validation	Self-censorship	Independent testing and verification
Browser testing	`/browse` Real Chromium Automation	none	Built-in browser automation
Teamwork	pass `.claude/skills/` Shared configuration	Personal use	Enterprise level, pay-per-seat

Use cases for gstack

Technology FounderThe CEO needs to balance his responsibilities with maintaining code output, and achieve the development efficiency of one person equaling twenty people through a structured process.
First-time users of Claude CodeProvides predefined expert roles and workflows to avoid being at a loss when faced with blank prompts.
Tech Lead and Staff EngineerIt requires rigorous code review, QA, and automated release processes to ensure that each PR undergoes multiple layers of review.
Parallel multi-project developmentIt is necessary to advance 10-15 sprints simultaneously, and manage multiple AI sessions through standardized processes to avoid chaos.
Teams that prioritize code qualityMechanisms such as dual AI cross-review, real browser testing, and automated regression testing are needed to ensure production safety.