I'm about to share my experience with a local, open-source AI alternative to Claude Code, and it's a wild ride!
Can a Free AI Stack Rival Claude Code?
I recently explored Goose and Qwen3-coder, two tools that, according to the internet, can combine to create a fully free Claude Code competitor. But is it really possible? I was determined to find out.
Setting Up the AI Trio
First, I downloaded Goose and Ollama. Then, I installed Qwen3-coder through Ollama. It was a straightforward process, but I learned a valuable lesson: always download and set up Ollama first! Otherwise, you might end up like me, trying to make Goose talk to Ollama without having Ollama ready.
Getting Ollama and Qwen3-Coder Running
I recommend installing Ollama first. I used the app version, which offers a user-friendly interface. Once installed, I chose the Qwen3-coder:30b model, which has about 30 billion parameters. This model is a whopper, weighing in at 17GB, so make sure you have ample storage space.
One of the key advantages of this setup is that your AI runs locally on your machine. No cloud, no data sent off-site. It's all right there on your computer.
Installing Goose: The Agent Framework
Next, it was time to install Goose. I chose the MacOS Apple Silicon Desktop version. When you launch Goose, you'll see a welcome screen with various configuration options. Since we're aiming for a free setup, scroll down to the "Other Providers" section and click "Go to Provider Settings."
Here, you'll find a long list of agent tools and LLMs. Scroll down, locate Ollama, and hit "Configure." This step is crucial as it sets up the connection between Goose and Ollama.
Configuring the Connection
You'll be prompted to choose a model. Again, select qwen3-coder:30b. Once you've configured both Ollama and the model, hit "Select Model," and voila! You've successfully installed and configured a local coding agent on your computer.
Testing Goose: The First Impressions
I decided to put Goose to the test with my standard coding challenge: building a simple WordPress plugin. Unfortunately, Goose/Qwen3 failed on the first attempt. It generated a plugin, but it didn't function as intended. Even after explaining the issue and trying again, it failed twice more.
By the third try, Goose managed to run the randomization but didn't fully adhere to the directions. It took five rounds for Goose to get it right, and it was quite pleased with itself, expecting perfection.
My Take on the Experience
I was a bit disappointed that it took Goose five attempts to succeed in my simple test. When I tested other free chatbots, most of them, except for Grok and an older version of Gemini, got it right on the first try. However, the difference with agentic coding tools like Claude Code and Goose is that they work directly on the source code, so repeated corrections can improve the codebase.
My colleague, Tiernan Ray, tried Ollama on a 16GB M1 Mac and found the performance unbearable. In contrast, I'm running this setup on an M4 Max Mac Studio with 128GB of RAM, and I've had multiple heavy-duty applications open simultaneously without issues. So far, the overall performance has been quite good, and I haven't noticed a significant difference in turnaround time compared to cloud-based products like Claude Code or OpenAI Codex.
Final Thoughts and Future Plans
These are just my initial impressions. To truly evaluate if this free solution can replace expensive alternatives like Claude Code's Max plan or OpenAI's Pro plan, I need to run a more extensive project through it. That's next on my agenda, so stay tuned for the full analysis.
Have you experimented with running coding-focused LLMs locally using tools like Goose, Ollama, or Qwen? What was your setup experience like, and what hardware did you use? If you've used cloud options like Claude or OpenAI Codex, how does local performance and output quality compare? I'd love to hear your thoughts in the comments below!
