Garry Tan Releases gstack: An Open-Source Claude Code System for Editing, Code Review, QA, and Deployment

What if AI-assisted coding became more reliable by separating product planning, developer reviews, releases, and QA into separate workflows? That’s the idea behind Garry Tan gstacka packable open source toolkit Claude Code in 8 workflow capabilities with views supported by the browser’s continuous runtime. The takeit describes itself as ‘Eight workflow skills with Claude Code concepts‘ and combines common software delivery activities into unique methods such as planning, updating, deployment, automated browser, QA testing, and retrospectives. The goal is not to replace the Claude Code with a new model layer. To make Claude Code work with clear role boundaries during product planning, engineering review, release, and testing.

8 Core Commands

I gstack The repository currently features 8 main commands: /plan-ceo-review, /plan-eng-review, /review, /ship, /browse, /qa, /setup-browser-cookiesagain /retro. Each command is mapped to a specific operating mode. /plan-ceo-review it is positioned as a product level planning pass. /plan-eng-review used for design, data flow, failure modes, and testing. /review focuses on production risk and code reviews. /ship it is used to configure the correct branch, sync with master, functional testing, and open PR. /browse gives the agent browser access, while /qa designed for systematic evaluation of affected routes and flows. /setup-browser-cookies it imports cookies from the local browser into the headless session, too /retro used for engineering retrospectives.

Continuous Browser is the Main Program

The most important technical part of gstack is not the Markdown capabilities. Browser subsystem. gstack provides Claude Code persistent browser and that the browser is the hard part, and the rest is mostly Markdown. Instead of launching a new browser for every action, gstack uses a a long-lived headless Chromium daemon and communicate with it localhost HTTP. The reason is the delay and maintenance of the state. A cold start costs approx 3–5 seconds per tool callwhile subsequent calls after initialization are designed to work approx 100–200 ms. Because the browser is always alive, cookies, tabs, localStorageand the login status persists for all commands. The server shuts down automatically after that 30 minutes of idle time.

How gstack Connects Browser Automation to QA

Those daemon structures are important for QA and browser-driven development. For many agent functions, browser automation is a separate step for debugging or a screenshot utility. In gstack, browser access is part of the core workflow. The repo is self-explanatory /browse such as a mode that allows an agent to log in, click on an application, take screenshots, and check for breakage. /qa build on that by analyzing the branch diff, identifying affected routes, and testing relevant pages or flows. A sample flow in the repo is shown /qa to test 8 files have been changed again 3 routes affectedand then checks those routes against the local application instance. This means that the project tries to integrate source changes into the actual behavior of the application instead of treating QA as a separate pass.

Installation Requirements and Project Structure

The cache launch options are also well specified. gstack requires Claude Code, Gitagain Bun v1.0+. I package.json shows the current version as 0.3.3list Playwright again diff as a runtime dependency, and includes a browse which can be made from browse fountain tree. According to the repo’s README, /browse includes native binaries and is supported in macOS and Linuxboth x64 again arm64. The install flow copies the repo ~/.claude/skills/gstackhe runs ./setupand enlists the skills of Claude Code. Teams can also copy the same setup to a local repository .claude/skills/gstack directory so workflows can be shared within a project.

Why Project Uses Bun

The architecture document describes what the project uses The Bun instead of the normal Node.js setup. There are 4 reasons mentioned: integrated binaries, native SQLite access, native TypeScript execution, and a built-in HTTP server Bun.serve(). Those decisions are more practical than cosmetic. gstack reads Chromium’s SQLite cookie database directly, and Bun’s built-in database support removes the need for additional native packages. The integrated binary model also fits the repo installation style, because users are not expected to carry a separate runtime toolchain internally. ~/.claude/skills/.

Key Takeaways

• gstack is a workflow layer for Claude Code, not a new model or agent framework. It packs software delivery into 8 capabilities with slash-command concepts for editing, updating, deploying, default browser, QA, cookie setting, and lookback.
• A persistent browser daemon is a core technology component. gstack uses a long-lived headless Chromium process over local HTTP so that cookies, tabs, localStorageand the login status persists for all commands.
• QA is tied directly to code changes. I /qa the workflow analyzes branch variations, identifies affected routes, and tests the appropriate application methods instead of handling browser testing as a separate manual step.
• The project was built around the Bun for practical reasons. Bun is used for integrated binaries, native SQLite access, native TypeScript execution, and a built-in HTTP server for the browser daemon.
• gstack’s contribution is a functional architecture. Its main value is to separate product reviews, engineering reviews, code reviews, releases, and browser-driven validation into transparent processes with minimal responsibilities.

Check it out Repo here. Also, feel free to follow us Twitter and don’t forget to join our 120k+ ML SubReddit and Subscribe to Our newspaper. Wait! are you on telegram? now you can join us on telegram too.