Getting Codex Working with Ollama

[daveManDaveDude]

As an experiment I am trying to use a local LLM namely qwen:14b and deepseek-r1-tool-calling:14b running against Ollama. I run codex in an activated python venv and I give a simple prompt such as "make me a hello world space invaders game in python". Codex thinks and then comes back with a workable solution but stalls when it trys to call a tool to patch the file it’s trying to create. When it stalls codex returns the prompt back to the user and displays a lump of json with the patch details in it. The output looks slightly different between DeepSeek and Ollama. Qwen output Abbreviated for brevity below.

Is there anything I can do to address this? I am guessing that codex is not supporting the tool calls that Ollama is returning.

    Yes, that should cover the basics. Let me put this all together in the code file.
    </think>

    <tool_call>
    {"name": "apply_patch", "arguments": {"cmd": ["apply_patch", "*** Begin Patch\n*** Create File: 
    space_invaders.py\n```python\nimport pygame\nimport sys\n\n# Initialize Pygame\npygame.init()\n\n# Set up the 
  ...
    </tool_call>

[daveManDaveDude]

I have answered my own question. I had gone directly from using the default openAI model to a local model on my Mac mini. My expectations were a little off :-) I Eventually I managed to get a local model to write a 1 line hello world in python after a lot of thinking, and trying to work out how to write a file.

So my answer to myself is Codex CLI is fine, you just need a much more powerful model!

Thanks so much to the entire team and community for codex and all things open source