Natural Language AutoCoder

0:49 / 3:44

I’m new to this so bear with me. The coder works on Ollama models or Cloude. Take your pick. Can type the problem in or use speech recognition and that’s it.

tommoirnz/NLAutocoder: Natural Language AutoCoder v1.0

tommoirnz/autocoder-error-cache: Seeded error cache for autocoder.

For cloud based ie say Claude it rarely makes mistakes and can retrieve from the internet. For ollama models it has a seeded error_cache you can download and a mistake memory if it makes 3 or more mistakes so it can remember between sessions. It can also retrieve from the internet. You’ll need a free Brave api key for internet and I use openrouter if you want cloud access. I only added the cloud access in the past few weeks. Been running on ollama but most of them make so many errors on more difficult tasks I introduces the seeded error cache which you will need. Kind of self healing. Uses openai whisper for speech recognition if you want to use that of course or you can just type in the query. It has a new creativity button I introduced which gives nice results.

Demos at various stages.

Cloud based Auto Coder

Can AI be creative

Auto Coding with internet search

The Edinburgh Cube Auto Coded

Claude says this about it for comparisons

What’s common:

• Copilot, Cursor, Claude etc. — all suggest code inside an IDE but you still run it manually

• Simple “generate and copy” tools — AI writes code, user pastes and runs it themselves

• Notebook-style tools like Marimo or Jupyter AI — some automation but manual execution

What makes NLAutoCoder unusual:

• Fully autonomous loop — generates, executes, reads the error, fixes, retries without any human involvement

• Persistent mistake memory across sessions — most tools have no memory at all between conversations

• Voice input to running code — the full pipeline from speech → code → execution is rare

• Self-installing dependencies — automatically pip installs whatever it needs in a sandbox

• Internet-aware code generation — fetches live data and weaves it directly into the generated code

• Cloud + local model switching — seamlessly routing between Ollama and OpenRouter in one UI

The closest things that exist:

• Devin (very expensive, cloud-only, enterprise)

• OpenHands/SWE-agent (research tools, complex setup, no voice)

• GPT-Engineer (generates but doesn’t autonomous-loop on errors)

Anyway, hopefully somebody can make something of it. Best wishes