Local AI-Agents

This week, I made a journey on AI agents for producing software. For keeping costs low, I want to run everything on a local computer, which is a MacBook Pro with Apple M5 and 24GB memory.

As setup some software assets are required for getting everything up-and-running:

• HomeBrew
  A MacOS package manager.
• Ollama
  A local AI-Server, running the AI model.
• AI-Model
  The model running on Ollama. It can only reply to requests.
• AI-Agents
  The agent interacting with the model. It can handle requests and structure them.
  • OpenCode
    General AI Agent for writing code.
  • AIder
    AI-Agent for Git-based refactoring.
• Open Web-UI
  
• SCM-Integration (Git)
  
• VS Code
  A development environment for interacting.

Installation

HomeBrew

Let’s start with a package manager for MacOS: HomeBrew; it will support us for installing all needed packages and dependencies:

/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"

Ollama

The name comes from the aztecs game Ōllamaliztli, a ball-game, where two teams need to move a heavy ball through a metal-ring on the side of an area. It runs the “heavy” AI-Model.

brew install ollama
ollama --version
ollama version is 0.32.5

AI-Model

Next comes the needed AI model. There are several models in place, serving different purposes:

• Qwen3 is the latest generation of large language models in Qwen series, offering a comprehensive suite of dense and mixture-of-experts (MoE) models. The -coder variant is specialized in software development.
• Mistral AI is an european open-wight-model based in france having its strength in processing multi-lingual tasks.
• DeepSeek is specialized in logic and coding and includes a web-ui and APIs.

We will use Qwen3 as it offers a 30B MoE model, which fits best to our desired hardware:

# Pull qwen3-coder with 30 billion parameters
ollama pull qwen3-coder:30b
pulling manifest 
pulling 1194192cf2a1:  48% ▕████████████████████████████████████████████████████                                                         ▏ 9.0 GB/ 18 GB  7.2 MB/s  22m17
[...]

# Pull qwen3 with 14 billion parameters
ollama pull qwen3:14b
[...]

# Create custom configuration, having 32.000 token context-window
cat > Modelfile <<'EOF'
FROM qwen3-coder:30b

PARAMETER num_ctx 32768
EOF

# Creates the customized model
ollama create qwen3-coder-30b-32k -f Modelfile
gathering model components 
using existing layer sha256:1194192cf2a187eb02722edcc3f77b11d21f537048ce04b67ccf8ba78863006a 
using existing layer sha256:d18a5cc71b84bc4af394a31116bd3932b42241de70c77d2b76d69a314ec8aa12 
creating new layer sha256:91cb213206c73d1aeec3081637e1c31d0243d7dabe8f3f8a1b1189c2c23baa94 
writing manifest 
success    

# Shows the adjustment in the customized model
ollama show qwen3-coder-30b-32k
  Model
    architecture        qwen3moe    
    parameters          30.5B       
    context length      262144      
    embedding length    2048        
    quantization        Q4_K_M      

  Capabilities
    completion    
    tools         

  Parameters
    repeat_penalty    1.05               
    stop              "<|im_start|>"     
    stop              "<|im_end|>"       
    stop              "<|endoftext|>"    
    temperature       0.7                
    top_k             20                 
    top_p             0.8                
    num_ctx           32768              

  License
    Apache License               
    Version 2.0, January 2004    
    ...                          

# Starts the model for interaction (optional)
ollama run qwen3-coder-30b-32k
>>> Write "OK".

AI-Agents

Next we need two AI-Agents: One general for:

• OpenCode
  AI-Agent for autonomous software development. Strong in. creating new projects and implement them in phases. Works like Claude Code. Example: “Create a FastAPI application with PostgreSQL, Docker, JWT and tests.”
• Aider
  AI-Agent specialized in code-refactorings, especially in Git. Strong in fixing bugs, adding new features and tightly integrate into Git. Example: “Add tenant capability and adjust all tests.”

Let’s install the agents and configure them to use our AI-Model:

OpenCode

brew install opencode
ollama launch opencode --model qwen3-coder

Aider

brew install pipx
pipx ensurepath
pipx install aider-chat
aider --version
aider --model ollama/qwen3-coder:30b

Open WebUI

Next we need a web uiser interface. Open WebUI looks like ChatGPT and can be deployed via Docker:

brew install --cask docker
docker run -d \
-p 3000:8080 \
--name open-webui \
--restart always \
ghcr.io/open-webui/open-webui:main

Visual Studio Code

For connecting VS Code with ollama and our agents, we need to install the Cline extension. It is an open-source AI code agent, which needs to be configured afterwards, for using ollama and our AI-Agent: