Local AI To-Do App with Ollama: Private and Offline

Why run the AI locally?

Three reasons people choose a local model for planning:

1. Privacy. The text of your tasks never goes to a cloud AI provider.
2. Cost. After you have the hardware, each request is free.
3. Independence. It keeps working without an internet connection to a provider, and no provider can change terms under you.

The cost is quality and speed: local models are weaker than the best cloud models and need a capable computer.

What you need

• A computer with enough RAM for the model. As a rough guide, 8-billion-parameter models run on a modern laptop with 16 GB of memory; check the model's page for exact requirements.
• Ollama, a free app that runs open models locally.
• DonDon on web or desktop (the OpenAI-compatible provider works in both).

Step 1: install Ollama and pull a model

Install Ollama from its website, then in a terminal:

ollama pull qwen3:8b
ollama run qwen3:8b "Say hello"

If the model answers, Ollama is working. Ollama serves an OpenAI-compatible API at http://localhost:11434/v1.

Step 2: connect it to DonDon

1. Open Settings → AI.
2. Choose the OpenAI-compatible provider.
3. The base URL defaults to http://localhost:11434/v1; leave it.
4. Enter the model name exactly as you pulled it, for example qwen3:8b.
5. No key is needed for a local endpoint (it is optional).
6. Turn AI on.

Nothing is sent anywhere while AI is off.

If DonDon runs in the browser, your browser must be allowed to call the local endpoint. If requests fail, check Ollama's CORS setting (the OLLAMA_ORIGINS environment variable) or use the desktop app.

Step 3: try it on real tasks

• Quick add: type a messy sentence and let the model structure it.
• Break down: paste a big task and ask for steps.
• Notes to tasks: extract action items from a meeting note.
• Weekly review: uses only numbers (counts and minutes) and dates, never task titles.

What to expect from small local models

Task	8B local model	Frontier cloud model
Parse "next Friday 9am"	Usually fine, occasionally off by a day	Reliable
Break a task into steps	Good	Very good
Schedule a full day with constraints	Hit and miss	Strong

Tip: keep prompts short, use a larger local model if you can, and review every proposal before accepting.

Hybrid option

Use the local model for everyday capture and switch to a cloud key only when you want a stronger plan. In DonDon you can change provider in settings at any time. See BYOK explained.

Limits to be honest about

• DonDon's calendar and time blocking, which auto-schedule suggestions rely on, are Premium features.
• There is no team or shared-list feature today.
• Local models vary widely; test before you rely on one.

Choosing a local model

Local models are measured in parameters (the "8B" in qwen3:8b means eight billion). Bigger usually means better and slower, and needs more memory.

Size	Typical hardware	What it handles well
3B-4B	Older laptops, 8 GB RAM	Parsing sentences, simple summaries
7B-9B	Modern laptops, 16 GB RAM	Task breakdown, short plans
14B-32B	Desktop or workstation with more RAM or a GPU	Scheduling with constraints

These are rough guides. Check each model's page for exact memory requirements, because quantisation and context length change the numbers. Start with an 8B model; move up only if the results are not good enough.

Tips for choosing:

• Prefer instruction-tuned models. They follow "return JSON" or "answer in a table" better.
• Mind the context window. Planning with many tasks needs room; very small contexts truncate your list.
• Try two or three on the same real tasks before settling.

Speed expectations

Local inference depends heavily on your hardware. A laptop without a dedicated GPU may take several seconds to a minute for a plan. That is fine for a once-a-day plan and annoying for instant quick-add. A common compromise is to use classic natural-language quick-add (no AI) for capture, and the local model for breakdown and weekly review, where a short wait is acceptable.

Troubleshooting

"Connection refused" or no response. Check that Ollama is running (ollama list should work) and that the base URL ends in /v1.

The model name is rejected. Use the exact name from ollama list, including the tag, for example qwen3:8b.

Works in the desktop app but not in the browser. Browsers enforce cross-origin rules. Allow the site's origin in Ollama (the OLLAMA_ORIGINS environment variable) and restart it, or use the desktop app.

Answers are poor or cut off. Try a larger model, shorten your input, or reduce the number of tasks per request.

It is slow. Use a smaller model, close other heavy apps, or run AI only for the tasks that need it.

Other local servers

Anything that exposes an OpenAI-compatible API works with DonDon's compatible provider: LM Studio, llama.cpp servers and similar tools. The base URL and model name are the only things that change. Hosted OpenAI-compatible services such as OpenRouter or Groq also work with the same provider setting and a key, but those are cloud services, so they are not offline.

What stays local and what does not

A local model keeps the AI request on your machine. It does not make the whole app offline:

• Your tasks, habits and notes still sync with DonDon's servers like any cloud app.
• Logging in, sync and billing still need the internet.
• The AI step itself works without a connection to any AI provider.

If you want a fully offline notes-and-tasks tool, a local-only app is a different category. DonDon's offline sync on mobile lets you keep working and sync later, but accounts and sync remain cloud-based.

A privacy-minded setup

1. Run a local model for everything that touches personal details.
2. Keep AI off by default and enable it only when you need it.
3. Use a cloud key only for non-sensitive planning, with a spending limit.
4. Review draft tasks before saving, because models make mistakes.
5. Keep your operating system and Ollama updated.

For comparisons with cloud AI, see best AI task manager apps.

Get started

Download DonDon from the platforms page and compare plans on pricing.

Frequently asked questions

Can DonDon use Ollama?

Yes. In AI settings choose the OpenAI-compatible provider. The default address is Ollama's local endpoint (http://localhost:11434/v1) and you enter the model name you pulled.

Does my task data leave my computer with a local model?

The AI request goes to the model running on your machine. Your tasks still sync to DonDon's servers as usual; only the AI step stays local.

Which model should I start with?

A small instruction-tuned model such as qwen3:8b or llama3.1:8b is a good starting point on a laptop with enough memory. Larger models give better planning but need more RAM.

Create a free account and start managing tasks today.

Smart task management — focus on what matters.

Create free account