Testing a Free, Open-Source Claude Code Rival: My Experience

In the rapidly evolving landscape of AI tools, two newcomers—Goose and Qwen3-coder—have emerged as potential free alternatives to Claude Code. Developed by Jack Dorsey’s company, Block, Goose serves as an open-source agent framework. In contrast, Qwen3-coder is a large language model (LLM) designed for coding tasks.

Introducing Goose and Qwen3-coder

• Goose: An open-source agent framework developed by Block.
• Qwen3-coder: A coding-centric LLM that allows for local development without cloud dependency.

Both tools promise to aid developers by enabling powerful local capabilities that rival paid options like Claude Code and OpenAI Codex. This article outlines my experience setting them up on a Mac, showcasing their initial performance and limitations.

Setup Process

To start, download Goose and Ollama, the LLM server needed for Qwen3-coder. During setup, I initially struggled as I had installed Goose before Ollama. The correct sequence is to install Ollama first, followed by the Qwen3-coder model, which integrates through Ollama.

Ollama can be downloaded as an application or via command line. For ease of use, I opted for the app version. Following installation, I selected the Qwen3-coder:30b model, a coding-optimized version with 30 billion parameters.

Installation Steps

• Install Ollama: Download and open the installer to select the coding model.
• Install Goose: After Ollama is set up, download Goose and configure it to connect with Ollama.

During configuration, I found it helpful to specify the directory from which Goose would operate. The initial test involved creating a simple WordPress plugin, which unfortunately encountered several issues.

Testing the Tools

My first attempt to generate a functional plugin with Goose and Qwen3-coder resulted in a failure. Subsequent trials required multiple corrections before the tool finally succeeded.

• First Attempt: Failed to generate a functional plugin.
• Second and Third Attempts: Required adjustments due to errors.
• Final Attempt: Successful after multiple corrections.

Performance Insights

Initially, I was disappointed that it took five tries for Goose to produce a satisfactory plugin. In comparison, other free chatbots typically perform better with fewer attempts. However, the iterative improvements with agentic coding tools like Goose suggest potential for refinement.

While running tests on a high-spec M4 Max Mac Studio with 128GB of RAM, the performance remained consistent despite running multiple applications simultaneously.

Future Considerations

In conclusion, while my tests with Goose and Qwen3-coder show promise, more extensive projects will better determine their effectiveness compared to paid options like Claude Code and OpenAI Codex. Future analyses will explore their capabilities further.

If you have tried integrating these tools into your coding workflow, share your experiences and insights. Your feedback can help guide others considering the transition to free AI coding solutions like Goose and Qwen3-coder.