LOCAL CODING AI

Use local AI for coding without pretending it is magic

Local models can be useful coding assistants, particularly for explanation, small functions, tests and private code. The gap between “helpful” and “safe to merge” is still your review and test suite.

Good starting tasks

Ask for an explanation of a function, generation of unit tests, a small pure function, repetitive type changes or suggestions for simplifying code. These tasks make it easier to verify whether the model understood the problem.

Repository context is the hard part

A coding model does not automatically understand your whole project. Editors and agents may select files, use retrieval or expose tools. Larger repository context increases token use and can exceed the model's context window.

Private code can stay local

If the editor, model runtime and any retrieval or tool integrations are configured locally, source code can remain on your own machine. Remote MCP servers, hosted embeddings and cloud fallbacks change that boundary.

Do not judge only by model size

A larger general model is not guaranteed to outperform a smaller model on your exact language or framework. Use the same small task and test suite when comparing candidates.

Run the code

Generated code can compile incorrectly, misuse an API or silently change behaviour. Treat suggestions as untrusted patches until linting, type checks and tests pass.

Use the task finder for a memory-filtered shortlist, then compare model requirements in the model directory.