Goose and Quen 3 Coder: Building Web Apps in Under 5 Minutes

A recent tweet from Jack Dorsey sparked considerable interest: "Goose + Quen 3 coder = wow interesting." Intrigued, I decided to put this combination to the test by building two distinct web applications from the ground up. The outcomes were nothing short of remarkable.

What is Goose?

For those unfamiliar, Goose (official name: codename goose) is an open-source local AI agent designed to automate tasks and generate code. It's important not to confuse it with Goose AI, a completely different project. Developed by Jack Dorsey's company, Block, Goose aims to compete with major players like Claude, Code, or Cursor.

Goose offers flexibility, running as both a CLI tool and a desktop client. A key advantage is its cost-free nature and the ability to connect with your preferred AI provider. You can opt for commercial services such as Anthropic, OpenAI, or Gemini, or run it locally using Olama with your choice of AI models.

However, this article will focus on the synergy between Goose and Quen 3 Coder. For the uninitiated, Quen 3 Coder is currently one of the most popular coding models available. For a detailed analysis, numerous excellent deep dives on Quen 3 Coder are available online.

Getting Started with Goose

While Goose can be installed and run from any terminal, this exploration utilizes its user-friendly desktop client. Upon launching the application, you're greeted by a clean interface. The initial step is to navigate to the settings and configure your AI provider. From there, you can select a specific model from that provider. If your desired model isn't listed, you can manually input its name. For this experiment, I used Open Router with the Quen 3 Coder model.

Project 1: Rock, Paper, Scissors with a Twist

With the setup complete, I initiated a new project. The first goal was to build a simple rock-paper-scissors game where the player interacts with the computer by making hand gestures recognized by a webcam. I provided only the core instructions, leaving the choice of tools and languages entirely up to the agent.

In just about 3 minutes, Goose completed the task with an incredible result. It autonomously chose to implement the MediaPipe library along with a hands detection library for gesture recognition, and the execution was flawless. The system accurately distinguished between rock, paper, and scissors, and the gameplay against the computer was seamless.

Next, I requested a few enhancements: adding engaging animations upon game completion and integrating SQLite to store player names and game history locally. This second iteration required more time, as Goose had to refactor the project to incorporate a Node.js server and the database connection. Nevertheless, once finished, the application worked perfectly without any need for manual debugging.

The result included confetti animations for a win, falling sad emojis for a loss, and a sword-clash animation for a draw. Furthermore, the game log was successfully recorded in the SQLite database. I confirmed this by refreshing the application; the log persisted. Entering a new name correctly generated a new player profile.

Remarkably, the total cost for building the entire project, as reported on the Open Router dashboard, was a mere 19 cents.

Code Quality and Structure

A review of the project in VS Code revealed well-structured and high-quality code. The project's architecture was logical, the Node.js Express server was correctly configured, and the SQLite integration was seamless. The front-end was straightforward, built with vanilla JavaScript, which was expected since I hadn't specified any particular frameworks.

Customizing Goose with 'Hints'

This raises an important question: can you guide Goose to use specific languages, frameworks, or coding conventions? The answer is yes, through a feature called 'hints.' A hints file allows you to define project structure, preferred tools, and other key parameters. This file can be applied on a per-project basis or globally to all your projects.

Project 2: A Dog Photo App with Specific Tech Stack

I put this to the test in my next project. I created a dedicated hints file specifying the use of React, TypeScript, and Zustand for state management, along with commands for the desired project structure. The prompt was to build a simple application that fetches random dog pictures and allows users to save them to a favorites list—a perfect test case for the specified tech stack.

Goose generated the project in approximately two minutes. However, I encountered some initial issues. While functionally perfect, the application's visual styling was off. Based on previous tests, it seems Goose sometimes struggles with setting up Tailwind CSS, which it appears to favor by default even when not instructed. After prompting it to use simple CSS instead, the subsequent iteration was visually appealing and polished.

The final app worked flawlessly and looked great. I could fetch random dog pictures, add them to a favorites list, manage that list, and all data was persisted locally. The code structure was clean, and it included a helpful README file with setup instructions. As requested in the hints file, it correctly used Zustand for state management.

A minor critique is that the API calls were not placed in a separate api folder as specified in the hints file, but this was a small detail in an otherwise successful build.

Final Thoughts

Overall, the experience of using Goose with the Quen 3 Coder model was genuinely impressive. The rock-paper-scissors game, in particular, was a standout success. It understood the instructions perfectly from the start, requiring no bug fixes or troubleshooting.

Goose is a compelling tool. It features an intuitive interface and highly customizable options for models and providers. It also offers numerous other powerful features like extensions, recipes, and a headless mode that weren't explored in this article. It has certainly earned a place in my development workflow.

There are very few downsides to this tool. Block has done an excellent job creating a powerful and user-friendly open-source AI agent.