What if AI-assisted coding grew to become extra dependable by separating product planning, engineering overview, launch, and QA into distinct working modes? That’s the concept behind Garry Tan’s gstack, an open-source toolkit that packages Claude Code into 8 opinionated workflow expertise backed by a persistent browser runtime. The tookit describes itself as ‘Eight opinionated workflow expertise for Claude Code‘ and teams widespread software program supply duties into distinct modes akin to planning, overview, transport, browser automation, QA testing, and retrospectives. The purpose is to not change Claude Code with a brand new mannequin layer. It’s to make Claude Code function with extra specific function boundaries throughout product planning, engineering overview, launch, and testing.
The 8 Core Instructions
The gstack repository at present exposes 8 foremost instructions: /plan-ceo-review, /plan-eng-review, /overview, /ship, /browse, /qa, /setup-browser-cookies, and /retro. Every command is mapped to a selected working mode. /plan-ceo-review is positioned as a product-level planning move. /plan-eng-review is used for structure, information stream, failure modes, and exams. /overview is concentrated on manufacturing danger and code overview. /ship is used for making ready a prepared department, syncing with foremost, operating exams, and opening a PR. /browse provides the agent browser entry, whereas /qa is designed for systematic testing of affected routes and flows. /setup-browser-cookies imports cookies from a neighborhood browser into the headless session, and /retro is used for engineering retrospectives.
The Persistent Browser Is the Core System
Crucial technical a part of gstack isn’t the Markdown expertise. It’s the browser subsystem. gstack provides Claude Code a persistent browser and that the browser is the arduous half, whereas the remaining is principally Markdown. As a substitute of launching a recent browser for each motion, gstack runs a long-lived headless Chromium daemon and communicates with it over localhost HTTP. The reason being latency and state retention. A chilly begin prices round 3–5 seconds per device name, whereas subsequent calls after startup are designed to run in roughly 100–200 ms. As a result of the browser stays alive, cookies, tabs, localStorage, and login state persist throughout instructions. The server additionally shuts down robotically after half-hour of idle time.
How gstack Connects Browser Automation to QA
That daemon structure issues for QA and browser-driven improvement. In lots of agent workflows, browser automation is a separate debugging step or a screenshot utility. In gstack, browser entry is a part of the core workflow. The repo describes /browse because the mode that lets the agent log in, click on by way of the app, take screenshots, and examine breakage. /qa builds on prime of that by analyzing the department diff, figuring out affected routes, and testing the related pages or flows. The pattern stream within the repo reveals /qa inspecting 8 modified recordsdata and 3 affected routes, then testing these routes towards a neighborhood app occasion. This implies the challenge is attempting to tie supply adjustments to precise software habits as a substitute of treating QA as a indifferent guide move.
Set up Necessities and Challenge Format
The repository’s implementation selections are additionally pretty particular. gstack requires Claude Code, Git, and Bun v1.0+. The package deal.json reveals the present model as 0.3.3, lists Playwright and diff as runtime dependencies, and compiles a browse executable from the browse supply tree. Based on the repo’s README, /browse compiles a local binary and is supported on macOS and Linux, for each x64 and arm64. The set up stream copies the repo into ~/.claude/expertise/gstack, runs ./setup, and registers the abilities for Claude Code. Groups can even copy the identical setup right into a repository-local .claude/expertise/gstack listing so the workflow is shared inside a challenge.
Why the Challenge Makes use of Bun
The structure doc explains why the challenge makes use of Bun fairly than a extra standard Node.js setup. There are 4 acknowledged causes: compiled binaries, native SQLite entry, native TypeScript execution, and a built-in HTTP server with Bun.serve(). These selections are sensible fairly than beauty. gstack reads Chromium’s SQLite cookie database immediately, and Bun’s built-in database help removes the necessity for additional native packages. The compiled binary mannequin additionally matches the repo’s set up model, as a result of customers will not be anticipated to handle a separate runtime toolchain inside ~/.claude/expertise/.
Key Takeaways
- gstack is a workflow layer for Claude Code, not a brand new mannequin or agent framework. It packages software program supply into 8 opinionated slash-command expertise for planning, overview, transport, browser automation, QA, cookie setup, and retrospectives.
- The persistent browser daemon is the primary technical part. gstack runs a long-lived headless Chromium course of over localhost HTTP so cookies, tabs,
localStorage, and login state persist throughout instructions. - QA is tied on to code adjustments. The
/qaworkflow analyzes department diffs, identifies affected routes, and exams the related software paths as a substitute of treating browser checks as a separate guide step. - The challenge is constructed round Bun for sensible techniques causes. Bun is used for compiled binaries, native SQLite entry, native TypeScript execution, and a built-in HTTP server for the browser daemon.
- gstack’s contribution is operational construction. Its foremost worth is separating product overview, engineering overview, code overview, launch, and browser-driven validation into specific modes with slender obligations.
Take a look at Repo here. Additionally, be happy to observe us on Twitter and don’t neglect to affix our 120k+ ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
