Local AI coding worker

I can hand a small coding job to a local model, then judge what comes back.

Format: Ollama / Codex / MCP
Status: WORKING EXPERIMENT

PROJECT MAP

CODEX
  bounded task
      ↓
LOCAL MODEL
  proposal
      ↓
CODEX → review → apply → test

The shape of the work.

01 / The project

What I’m working with

The designer wanted local hardware to do useful work alongside hosted Codex. The worker accepts a bounded task and returns a proposal.

I still have to review the proposal, apply useful changes and check the result. A well-formed answer can contain bad code.

02 / Workflow

How it works

Define a small task

Send the worker the specific context and output contract it needs.

Get a local proposal

The Ollama-backed MCP worker generates a response within that contract.

Review the result

Hosted Codex decides what to apply and verifies the resulting change.

03 / Inside the project

Inside the project

MCP client and worker contractBounded requests and returned proposals.
Local-model setupOllama server and Windows client setup.
Benchmarks and outcome recordsSynthetic task results, failure cases and rollout notes.

04 / Status

Where I leave it

The worker and MCP connection exist. Real project outcomes matter more than a synthetic benchmark score; those outcomes are part of the continuing experiment.