This is a fully working coding agent, much like Codex CLI. (but of course I’m sure far less capable, even if surprisingly so in its own way)
Built with Go so it doesn’t take a year to compile and yet performs super fast.
It works with OpenAI API and Ollama.
It can:
read, write, edit and run arbitrary bash commands
it presents code deltas and commands for you to review and approve
it maintains an evolving context summary beyond the last X interactions so it uses tokens more efficiently
it maintains a session for the current working directory
it is guardrailed to the current directory (best efforts!)
you can interrupt the agent and set it off in a different direction
I think the fun stat here is the number of lines of code, using CodeGollm as a baseline:
Codex repo is 250x bigger
Pi-mono coding agent (known for its small size) is 7x bigger
What I learnt doing this is:
The fundamentals of all this remain pretty straightforward.
You really can achieve an awful lot with a small harness
Codex CLI will happily build you something you can replace it with
Model size really matters - Gemma 4 was a perhaps predictable disaster, but gpt-4.1-mini can code pretty well, despite not racking up internal reasoning!
You probably won’t be using this because of course I can’t build something to compete with Codex CLI in a day and there is no ChatGPT auth, but I thought it interesting to share . Give it a on GitHub if you enjoyed this … have fun building!
Unfortunately it didn’t much , but this is still useful for any Ollama models and might work well with larger models.
gpt-5.4-mini with medium reasoning still shows off the capabilities of this coding agent well, however, and it’s very cost effective. The OpenAI models are just so awesome for this stuff, even mini
I’ve since updated the system prompt slightly though I’ve not properly evaluated that yet, but it definitely still works!
This was almost one-shotted apart from the tinted glass:
on GitHub if you enjoyed this … have fun building!
