Ollama (Local Models)

Run AI models locally on your machine — completely free, fully offline.

What Is Ollama?

Ollama is an app that lets you run open-source AI models locally on your computer. No API key, no usage costs, no internet required.

Install Ollama

1. Download Ollama from ollama.com/download
2. Install and open it (it runs in the background)
3. Codiko will guide you through downloading the right models during setup — no terminal commands needed

Configure in Codiko

On first launch

1. Choose Ollama as your AI provider in the setup wizard
2. The base URL defaults to http://localhost:11434 — change this if Ollama is running on another machine on your network (e.g. http://192.168.1.100:11434)
3. Click “Connect” — the app will verify the connection
4. Choose a model tier based on your hardware — the app will recommend one and download the models for you
5. Review the model assignments and click Next to continue

From the settings menu

1. Open Settings (gear icon, bottom-left)
2. Under AI Provider, select Ollama
3. The base URL defaults to http://localhost:11434 — change this if needed
4. Click “Test Connection” — the app will detect available models
5. Select models from the dropdowns

Hardware Requirements

Codiko will detect your hardware and recommend an appropriate model tier during setup. You can always change this later in Settings.

• 8–12 GB RAM: Lightweight tier (4B model) — basic assistance, less detailed responses and less reliable tool use, but still useful for brainstorming and simple tasks
• 12–24 GB RAM: Standard tier (9B model) — good balance of quality and speed, reliable tool use. Our recommended starting point
• 32 GB+ RAM: Performance tier (27B model) — best local quality, more detailed and nuanced responses, but slower inference
• Dedicated NVIDIA GPU: Significantly faster inference on any tier
• Apple Silicon Mac: Look for -mlx model variants (e.g. qwen3.5:9b-mlx) for better performance on M-series chips

Performance Note

Local models have a short loading delay before responses begin streaming — this is normal and depends on your hardware and the model size. Once streaming starts, you’ll see the model’s thinking process in real-time. For the fastest experience overall, cloud providers (OpenAI or Anthropic) are still the benchmark.

Advanced Settings

Fine-tune how Ollama works with your project in Settings → AI:

• Context Window Size — how much conversation and project context the model can see per message. Larger values give better results but use more memory. This is set automatically during setup based on your hardware but can be changed at any time.
• Story Bible Injection — automatically includes your full story bible in every message so the model has immediate access to character, location, and world-building data without needing to look it up. Recommended for smaller models (9B and below) that may not proactively search for this information

Recommended Models

We develop and test primarily with the Qwen 3.5 family. Other models may work with varying levels of tool-use reliability.

If you run into issues with a specific model, you can submit a bug report using the button [] at the top of the chat window. We’re always working on improving compatibility so please let us know if you find any issues.

The app uses three model roles. For the Chat Model, use a capable model that supports tool use (e.g. qwen3.5:9b or qwen3.5:27b depending on your hardware). For the Router and Summarise models, a smaller model like qwen3.5:4b works well.

If your machine struggles with larger local models, Ollama also offers cloud-hosted models (like Kimi K2, DeepSeek, and Gemma 4) that don’t require powerful hardware.

See Recommended Models for a full comparison and hardware guidelines.