Reviewing a Free, Open-Source Local Alternative to Claude Code
▼ Summary
– Jack Dorsey’s post sparked interest in the free, open-source AI coding tools Goose (an agent framework) and Qwen3-coder (a coding LLM), which together could rival paid services like Claude Code.
– Setting up the local AI coding environment requires installing Goose, Ollama (an LLM server), and the large (17GB) Qwen3-coder model, which runs entirely on your own machine without cloud dependency.
– The setup process is straightforward but demands a powerful computer with significant storage and memory, as performance on less capable hardware can be poor.
– Initial testing showed the Goose/Qwen3-coder combination succeeded on a coding task but required multiple retries and corrections to achieve a working result.
– While early performance on high-end hardware is promising and comparable to cloud-based alternatives, a definitive assessment of its ability to replace paid plans awaits more extensive project testing.
Exploring a free, open-source alternative to premium AI coding assistants like Claude Code can be an exciting venture for developers looking to keep their work local and private. The combination of Goose, an agent framework from Block, and Qwen3-coder, a specialized coding model, promises a fully offline coding assistant without subscription fees. This setup leverages your own hardware, ensuring data never leaves your machine, though it demands significant local resources to function smoothly.
The journey began with a cryptic social media post from Jack Dorsey, hinting at the potential of pairing Goose with Qwen3-coder. Intrigued, I decided to test whether this duo could genuinely compete with paid services. This initial piece walks through the installation and configuration process, with future articles planned to delve deeper into each component’s role and a practical build project.
Getting everything operational starts with downloading the necessary software. You’ll need both Goose and Ollama, with the Qwen3-coder model fetched later through Ollama itself. A word of advice: install Ollama first. I learned this the hard way after initially installing Goose and finding it couldn’t communicate with Ollama, which wasn’t yet present on my system.
For Ollama, I opted for the desktop application on macOS, though command-line versions are available for other operating systems. Upon launching, the interface presents a chat window. The default model listed wasn’t what I needed, so I selected Qwen3-coder:30b from the model list, the “30b” denotes its 30 billion parameters, tailored for coding tasks. The model, a substantial 17GB download, only initiates when you first submit a prompt, so ensure you have ample storage space.
A crucial step involves making the Ollama instance accessible to other local applications. This is done by enabling the “Expose Ollama to the network” option in the settings menu. I also adjusted the context length to 32K, a conservative setting given my machine’s 128GB of RAM, to observe performance within a constrained environment. Notably, creating an account or using cloud features was avoided to maintain a completely free, local setup.
With Ollama configured and running, attention turned to Goose. After downloading the appropriate version for my system, the initial launch presents a welcome screen. Navigating to the provider settings reveals a long list of compatible tools and models. Scrolling to find Ollama and selecting it begins the connection configuration. Here, you choose the specific model again, qwen3-coder:30b, to link Goose to your local LLM. This step essentially tells Goose how to talk to the model running via Ollama on your computer.
Once connected, you can direct Goose to a specific working directory and confirm the active model is displayed. For an initial test, I used my standard challenge: creating a basic WordPress plugin. The first attempt was unsuccessful; the generated code did not function. Subsequent tries, where I provided feedback on the errors, also fell short. It wasn’t until the fifth iteration that Goose, with Qwen3-coder, finally produced a working plugin, and it seemed quite satisfied with the result.
These early tests reveal both promise and room for improvement. The need for multiple retries was disappointing, especially when compared to several free chatbots that passed the same test on their first attempt. However, a key distinction exists: agentic tools like Goose and Claude Code operate directly on source code files, so iterative corrections actively refine the final codebase, which is a different workflow from simple chatbot interactions.
Performance on my well-equipped M4 Max Mac Studio with 128GB of RAM was quite responsive, even with numerous other demanding applications running. I didn’t perceive a noticeable delay compared to hybrid local/cloud services. It’s worth noting that a colleague testing on a machine with 16GB of RAM found the experience much less tolerable, underscoring that substantial local hardware is a prerequisite for this approach.
While these first impressions are encouraging, the true test of whether this free stack can replace plans costing $100 or $200 per month will come when applying it to a large, complex project. That evaluation is still ahead. For developers with powerful machines seeking privacy and cost savings, Goose paired with Qwen3-coder presents a compelling, if still maturing, alternative worth exploring.
(Source: ZDNET)
