Dethinker: Strip 'thinking' content from Ollama responses

I try to use Local LLMs whenever I can, pretty much always via Ollama. Recently with qwen3 I have been really impressed with the quality of the responses. However, I kept running into the problem that many applications don’t strip the “thinking” tokens in responses, which causes problems.

Rather than modifying each app individually (I did try a few 😭), I created a simple proxy server called Dethinker that sits between your client application and Ollama, automatically stripping out any “thinking” content from responses.

This works really well when using applications like HyprNote with Ollama, as I think we get better responses from qwen3 than llama3.2.

So What Does It Do

Dethinker is a lightweight proxy that:

• Sits between your client app and Ollama server
• Intercepts API responses both for Ollama and OpenAI-compatible APIs
• Strips out content between configurable thinking tags
• Passes the clean responses back to your application

How Do I Use It

All you need is uv installed. I am using the power of shebang and PEP 723 to have a self-contained executable.

git clone https://github.com/bhubbb/dethinker.git
cd dethinker
./dethinker.py

There is also a Dockerfile if you are into that…

Once running, point your client application to http://localhost:21434 instead of http://localhost:11434. That’s it.

By default, Dethinker looks for content between <think> and </think> tags, but you can configure different tags by setting environment variables before starting the server. Check the GitHub repository for configuration details.

Thanks for reading!