Hobby Coding AI Setup

Alright, I cannot go past it: AI is growing so big in coding that I cannot go past it. So instead of fighting, trying to understand it seemed a wiser option to me. So, why not applying it to my hobby coding?

Architecture

Requirements

A hobby setup cannot compare with a professional-grade setup: not only is the budget a key factor, it is also overkill for most of our needs. So here we go. The setup must:

  • Be installable and usable in both Windows and Linux environments
  • Be shared between multiple users in Windows environments
  • Support free open-weight models or support very low-cost models
  • Provide on-demand access to cutting-edge flagship models such as the latest Claude Code or Codex models. However, the default mode would rely on the free models
  • Have a CLI interface
  • Provide guardrails in case of mistakes with easy rollbacks

As a nice-to-have, the setup could:

  • Support local models or at least having a strong focus on privacy
  • Be accessible from VS Code

Plan

Ollama and OpenRouter are the cornerstone of the overall setup:

  • Ollama provides access to the free open-weight models (List of local models supported by Ollama). This will keep cost low and ensure privacy
  • OpenRouter provides access to flagship models without vendor lock-in. With a pay-as-you-go pricing, the setup can easily switch between Claude Code or Codex. While there is a platform fee (5.5% as of October 2026), it does not significantly impact the total pricing considering my low usage

Two interfaces will be added to interact with the models:

The considered models, as of October 2026, are:

Setup

Pre-requisites

On a Windows 11 machine, Node and Python need to be installed first.

Node.js

I don’t use many programming languages for my hobby coding. While I use ASDF at work, I will stick to a simple NVM for Windows setup.

Python

  • Install with winget:

    winget install Python.Python.3.12
    Using winget makes Python upgrade easier 
    # Check what you have installed
winget list Python

# See available upgrades (all winget packages)
winget upgrade

# Upgrade Python specifically
winget upgrade Python.Python.3.12

    Note: Moving between major versions (3.12 → 3.13) is NOT an upgrade — it’s a new install:

    winget install Python.Python.3.13

    Such major version upgrade requires a re-installation of Aider

Ollama

Installation

  • Check the Windows installation guide: https://docs.ollama.com/windows.

  • Install Ollama from https://ollama.com/download.

    I personally picked the manual installer so that I can install it into a shared location:

    ./OllamaSetup.exe /DIR="C:\Program Files\Ollama"

    This is required to have Ollama installed in a shared location across Windows users

    PowerShell must run in administrator mode for this.

  • Create a folder to store downloaded Ollama models. I created C:\Ollama\Models

  • Setup the Ollama models location to the newly created folder

  • Restart Ollama

Pull a model

  • Check the available local models: https://ollama.com/search?c=local

  • Ensure your device supports the model

    • As a reference, assume the entire model size dumped into VRAM (Video RAM). So a 4GB download will roughly consume 4GB from your GPU. The key metrics are the number of parameters involved. To read more: https://plugable.com/blogs/news/gpu-vram-requirements-for-local-llms-plugable-guide.

    • If your GPU is a Nvidia, check your VRAM usage with nvidia-smi. A useful display is:

      nvidia-smi --query-gpu=memory.total,memory.used,memory.free --format=csv

      With a qwen3.5 4B running, an example is:

      memory.total [MiB], memory.used [MiB], memory.free [MiB]
8151 MiB, 4626 MiB, 3266 MiB

    • Use the “Download size” filter from Ollama

  • Pull it:

    ollama pull qwen3.5:0.8b

  • Test the model

    • Run it

      ollama run qwen3.5:0.8b

      Type any prompt and a response should appear. If your device does not support it, an error will be thrown while attempting to run the model.

      Use the following command to monitor the currently active models:

      ollama ps

    • Open the Ollama Chat (Right click on the Ollama tray icon > Open Ollama > Select the “Chat” tab > pick the pulled model) and input a prompt

OpenRouter

  • Sign up for a new account https://openrouter.ai/
  • If necessary, top up your credit.
  • Open the dashboard (link) and create an API key

Cline extension

  • Install Cline from the extension menu in VS Code (link). Alternatively, install it from https://cline.bot/ide in your favourite IDE.
  • Open the Cline settings
  • Connect it with the models (https://docs.cline.bot/getting-started/authorizing-with-cline#ide-setup)
    • Ollama:
      • Select “Ollama” in API Provider
      • Select the model
    • OpenRouter:
      • Select “OpenRouter” in API Provider
      • Input your OpenRouter API Key
      • Select your desired model. Beware, costs may be incurred.
  • You are ready to go!
    • For additional safety, start with “Plan” mode with the bottom right toggle to avoid undesired changes

Aider

  • Check the installation guide and Ollama setup.

  • Install Aider:

    python -m pip install aider-install
aider-install

  • Configure the Ollama endpoint

    setx   OLLAMA_API_BASE http://127.0.0.1:11434

  • Restart the shell if needed

  • Move to your repository location

    cd <your_repository>

  • Start Aider

     aider --model ollama_chat/qwen3.5:4b --chat-language english

Workflow

The simple setup uses Cline + Ollama + Qwen family model, which is sufficient to revamp my personal website. However, the architecture was set up using Claude Code.

Aider is involved when multiple files are edited, such as a refactoring, or a non VS Code action (e.g. raw JSON file manipulation) is required.

As for OpenRouter, it is only used when flagship models are required.

Next steps

Now that the AI coding setup is completed, the Pandora box is opened: