I recently embarked on an exciting journey to explore the capabilities of local, open-source, and completely free AI coding tools, specifically Goose and Qwen3-coder. These tools, developed by Jack Dorsey's Block and Alibaba, respectively, have sparked curiosity in the AI community. The goal was to find a free alternative to Claude Code, a popular yet expensive coding assistant. But can these tools truly replace the premium plans? Let's dive in and find out.
The Setup:
The process began with downloading and installing Goose and Ollama, an LLM server. I opted for the MacOS Apple Silicon Desktop version of Goose, which offered a straightforward setup. However, I encountered a challenge when trying to connect Goose to Ollama, realizing the importance of a powerful local machine for this endeavor.
Model Selection:
After installing Ollama, I chose the Qwen3-coder:30b model, a coding-optimized LLM with 30 billion parameters. This model is a significant advantage, as it is local and doesn't require cloud access, ensuring privacy and potentially faster performance. I set the context length to 32K, considering my machine's RAM capacity.
Goose in Action:
With the tools installed, I tested Goose's coding abilities. I used a standard test challenge of building a simple WordPress plugin. Unfortunately, Goose/Qwen3 initially struggled, but after repeated attempts and corrections, it eventually succeeded. This process highlighted the importance of agentic coding tools working on actual source code, improving the codebase with each iteration.
Performance and Comparison:
Running this setup on my M4 Max Mac Studio with 128GB of RAM, I found the performance to be quite good. The turnaround time for prompts was comparable to local/cloud hybrid products like Claude Code and OpenAI Codex. However, I was disappointed that Goose required five attempts to complete the test, while other free chatbots achieved success on the first try. This discrepancy emphasizes the need for further refinement and optimization.
The Verdict:
While Goose and Qwen3-coder show promise, they are not yet ready to replace premium plans like Claude Code's Max plan or OpenAI's Pro plan. The setup process and performance are impressive, but the need for repeated corrections and the current limitations in accuracy suggest that more development is required. As an AI enthusiast, I'm excited to see the progress and potential of these open-source tools, and I look forward to future updates and improvements.
