We tested Gstack, the tool that promises to replace an entire team of developers

We tested Gstack, the tool that promises to replace an entire team of developers

The solution automates Claude Code to simulate a complete team of engineers. Its installation presents technical challenges for beginners. The tool successfully deploys a functional application.

On paper, the promise of Gstack, the open-source framework created by Garry Tan, CEO of Y Combinator and formerly of Palantir, is a dream. The tool should transform Claude Code into a virtual team of engineers. Several roles are assigned in this sense. “A CEO who rethinks the product, a technical manager who defines the architecture, a designer who corrects AI imperfections, a reviewer who detects bugs in production,” lists Garry Tan on his Github. But also “a quality assurance manager who opens a real browser, a security manager who carries out audits (…), and a deployment engineer who publishes the pull request”. In all, he summarizes, “twenty-three specialists and eight powerful tools, all in slash commands, all in Markdown, all free, under the MIT license.”

With Gstack, no more division of tasks between several solutions. To keep its commitment, the solution automates a Chromium browser. The goal: to allow Claude Code to test and manage end-to-end web applications. AI maintains continuous access to sessions and authentications. It relies on simplified benchmarks to navigate web interfaces.

To test the tool, we will create a mini-invoicing application. The opportunity to return to certain promises listed by Garry Tan to be taken with a grain of salt, especially for Windows beginners like us. We will see that, in practice, the experience is more nuanced. Gstack can save real time, but it does not remove the constraints of Windows, nor those of the browser, nor the usual limits of web development.

Step 1: installing Bun

According to Garry Tan’s Gstack repository section, the system requirements are “Claude Code, Git, Bun v1.0+, Node.js (Windows only).” The problem is that simply typing “node” in PowerShell puts you in interactive mode (REPL). In this one, the system commands no longer work, which can block many newbies. Hence a possible confusion for a beginner user between launching Node and launching a system command. On Windows, it is best to install Bun via npm (npm install -g bun). This is to avoid frustrating configuration issues.

In this sense, under PowerShell in Windows we type the following command:

npm install -g bun

To check that everything went well, we type: bun –version. If a number (like 1.3.12) appears, that’s great. This allows you to run certain project commands, install dependencies, and sometimes run repository-related scripts more simply.

Step 2: setting up Gstack

According to the GitHub repository, installation takes 30 seconds. “Open Claude Code and paste this. Claude will take care of the rest.” However, the ./setup command may fail on Windows due to a bug. To get around this, it may be necessary to compile each command manually, via Bun Build. What was supposed to take 30 seconds becomes longer than expected.

In order to carry out this installation on Windows, we open Git Bash. We create the necessary folder and download Gstack with these commands:

mkdir -p ~/.claude/skills git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/.claude/skills/gstack

We place ourselves in the Gstack folder:

cd ~/.claude/skills/gstack

We install the dependencies:

bun install

To dodge the known bug, we manually compile the Gstack tools by typing these four commands one after the other:

bun build src/cli/browse.ts --compile --outfile bin/browse.exe bun build src/cli/find-browse.ts --compile --outfile bin/find-browse.exe bun build src/cli/design.ts --compile --outfile bin/design.exe bun build src/cli/global-discover.ts --compile --outfile bin/gstack-global-discover.exe

This should make installing Gstack easier.

Step 3: Concept validation

The /office-hours command writes a design document. The Gstack repository refers to it as: “six essential questions to rethink your product before you even write a single line of code.”

The tool is excellent. But be careful, the Gstack README does not specify that if you close “Claude Code”, you must return to the correct folder (cd $HOMEinvoicing in our case) before relaunching it. Otherwise, the AI ​​loses practical anchoring in the correct working file.

For our part, in PowerShell, we created a folder for our project: mkdir $HOMEinvoicing We put ourselves in our folder before relaunching “Claude Code”: cd $HOMEinvoicing. Then we write: claude

In Claude, we type:

/office-hours "Je veux créer une mini-app de facturation en HTML/JS pur."

Claude asks us some pointed questions. For example: what legal notices? How to store data? How to generate the PDF? We respond naturally. At the end, he will write a clear design document.

Step 4: technical planning

Before building the house, you need the architect’s plans. The /plan-eng-review command acts as a technical manager. However, for a simple application, such as a single HTML file for invoicing, AI tends to offer complex architectures. For example, the Gstack README does not warn beginners that they have the right to refuse an architecture that is too heavy.

To avoid this, we simply type in Claude Code:

/plan-eng-review

This will read the previous design document and define the architecture. For example: how invoices are saved in the browser. It also lists edge cases, such as: what happens if the user enters letters in a price field? At this stage, we are careful not to always accept overly complicated architectural proposals.

Step 5: building the application

According to Garry Tan, Gstack can write a large number of lines of code, in different files, in a few minutes. It fails to discuss some technical limitations that can create time-consuming obstacles. For example, it is impossible to automatically attach a locally generated PDF to an email via a simple web button for security reasons. This mainly comes from a browser security constraint. The beginner can spend hours asking the AI ​​to do the impossible.

In our case, we simply ask Claude Code to build the application: “Implement the technical plan. Create the index.html file with all the necessary HTML, CSS and JavaScript code.”

Claude generates a very complete file with around 750 lines of code. To see the result, we open another PowerShell and type:

cd $HOMEfacturation

Then :

start index.html. 

Our browser opens and displays our application. We’re not trying to make it automatically attach a PDF to an email. We prefer to mark a message before choosing the mailbox so as not to forget it.

Step 6: automatic tests

According to Garry Tan, the /qa command acts as a quality tester. “Test your application, find bugs, fix them with atomic commits (i.e. an operation that applies a set of distinct changes in a single operation, editor’s note) then check again.”

What it doesn’t say is that Gstack’s QA system is very strict. The headless browser testing the interface may sometimes fail to click on elements that are functional for a human. An application may not receive full marks by having accessibility issues, focus issues, or unstable behavior.

So, for our part, we type in Claude Code: /qa

Then we agree not to reach the maximum /qa score given by Gstack before launching the application.

Step 7: upload

The /ship command acts as a deployment engineer. Be careful to identify GitHub to use it.

To do this, we type in PowerShell: gh auth login

We choose GitHub.com > HTTPS > Login with a web browser. We copy the code displayed on the screen and paste it on the GitHub web page to authenticate. We return to Claude Code and write: /ship

Claude automatically creates a code backup, an update request on GitHub and asks for some confirmations. A few minutes later, our application is publicly accessible via https://brunono251.github.io/facturation/

After a few bug fixes, the result is quite impressive. The application is functional. It allows you to easily create invoices. The design is clean. This allows you to generate an invoice in just a few clicks with no learning curve. Since the app is hosted on GitHub Pages, it loads instantly. The data is stored in the browser’s Local Storage.

Jake Thompson
Jake Thompson
Growing up in Seattle, I've always been intrigued by the ever-evolving digital landscape and its impacts on our world. With a background in computer science and business from MIT, I've spent the last decade working with tech companies and writing about technological advancements. I'm passionate about uncovering how innovation and digitalization are reshaping industries, and I feel privileged to share these insights through MeshedSociety.com.

Leave a Comment