Some things shouldn't leave your computer. Medical notes, legal drafts, journal entries, messages to your therapist, that honest performance review you're still editing. When you paste those into a cloud AI service, they travel to a server you don't control, get processed by infrastructure you can't inspect, and may end up in training data you never agreed to.
That bothered me enough to build local LLM support into WindowSill. And one of the easiest way to set it up is with Ollama, a free tool that runs AI models directly on your hardware.
This guide walks you through the setup. By the end, you'll have grammar checking, text rewriting, tone adjustment, and translation running locally in any Windows app, with zero data leaving your machine.
What you'll need
- WindowSill installed from the Microsoft Store (free tier is fine for setup, but AI features require WindowSill+)
- Ollama installed from ollama.com
- 8 GB of RAM minimum (16 GB recommended for comfortable use alongside other apps)
- A few gigabytes of disk space for the model you choose
No GPU required. Ollama runs on CPU too, just slower. If you have an NVIDIA GPU with 6+ GB of VRAM, responses will be noticeably faster.
Step 1: Install and start Ollama
Download Ollama from ollama.com and run the installer. Once installed, Ollama runs as a background service on your machine. You can verify it's working by opening a terminal and running:
ollama --version
If you see a version number, you're good.
Step 2: Pull a model that's good for writing
Not all models are equal for writing tasks. You want something that understands grammar, tone, and natural language well. Here are three solid options:
| Model | Size | Best for | Pull command |
|---|---|---|---|
| Deepseek R1 8B | ~8 GB | General writing, grammar, rewriting | ollama pull deepseek-r1:8b |
| Qwen 3.5 4B | ~4.5 GB | Lighter machines, still capable | ollama pull qwen3.5:4b |
| Deepseek R1 1.5B | ~1 GB | Fast responses, good grammar | ollama pull deepseek-r1:1.5b |
Open a terminal and pull whichever model fits your hardware:
ollama pull qwen3.5:4b
This downloads the model to your machine. It only needs to happen once.
A note on model size vs. quality: Larger models (13B+ parameters) produce better writing output, but they need more RAM and a decent GPU to run at a usable speed. For most writing tasks (grammar fixes, tone adjustments, short rewrites), an 8B model is more than enough. Start small and upgrade if you need to.
Step 3: Connect WindowSill to Ollama
Ollama exposes a local API at http://localhost:11434. WindowSill can connect to it automatically.
In WindowSill:
- Open the Settings from the command bar
- Go to the AI Writing & Analysis section
- Under AI Providers, select Ollama
- Set the Ollama access point to
http://localhost:11434 - Select the model you pulled (e.g.,
qwen3.5:4b)
That's it. WindowSill now routes AI requests to Ollama instead of a cloud service.
Step 4: Test it
Open any app where you write (Word, Outlook, Notion, Slack, a browser, anything). Type a sentence with a deliberate mistake:
Their going to the meeting tommorrow at 3pm, can you confirmed?
Select the text. WindowSill's Analyze / Rewrite sill should appear on the bar. Hit the Spell Check option.
If everything is connected, the corrected text will come back after a few seconds:
They're going to the meeting tomorrow at 3 PM. Can you confirm?
The first request might take a moment while Ollama loads the model into memory. Subsequent requests will be faster.
What you can do with local AI
Once connected, all of WindowSill's AI writing features work through your local model:
Grammar and spell check. Select text in any app, fix errors without opening a separate tool.
Rewriting. Highlight a paragraph and ask for a rewrite. Useful for polishing drafts or simplifying dense writing.
Tone adjustment. Switch between professional, casual, and attention-grabbing tones. You can also create custom tone presets for recurring needs (e.g., "customer support reply" or "executive summary").
Translation. Select text and translate to any of 35+ supported languages. The quality depends on the model you chose. Llama 3 handles common language pairs (English/Spanish, English/French, English/German) well. For less common pairs, a larger model or a specialized translation model works better.
Custom prompts. Build reusable prompts with variable injection. For example, a prompt that takes selected text and converts it into a formatted meeting recap with today's date auto-inserted.
Summarization. Select a long email or document section and get a summary.
All of this happens on your machine. Nothing goes to OpenAI, Google, or anyone else.
Performance tips
Local models are slower than cloud APIs. Here's how to keep things comfortable:
- Close the model when you're not using it. Ollama keeps models in memory. If you're done writing and need the RAM for something else, run
ollama stop qwen3.5:4bin a terminal. - Use a smaller model for quick tasks. Grammar checks and spell fixes don't need a large model. Qwen 3.5 4B handles these well and responds faster than Deepseek R1 8B.
- Use per-prompt model selection. WindowSill lets you assign different models to different tasks. Use a fast, small model for grammar and a larger one for complex rewrites. This is also useful if you want to keep local models for sensitive content but use a cloud model for non-sensitive tasks.
- Upgrade your RAM. If you have 8 GB, Ollama will work but your system will feel sluggish when the model is loaded. 16 GB gives you room to run Ollama alongside a browser, editor, and other apps without swapping to disk.
When local isn't enough
I want to be honest about the trade-offs. Local models are good, but they're not as capable as the latest cloud models for every task. Here's where I'd still use a cloud provider:
- Long, complex rewrites where nuance matters. GPT 5.5 and Claude 4.5 Sonnet still produce more polished output for multi-paragraph rewrites.
- Uncommon language translations. Local models handle major languages well, but for something like Finnish to Japanese, a cloud model with broader training data does better.
- Very fast response times. Cloud APIs respond in 1-2 seconds. Local models on CPU might take 5-15 seconds for a similar response.
The good news: you don't have to pick one. WindowSill supports both local and cloud providers at the same time. You can route sensitive content through Ollama and non-sensitive content through a cloud API. Per-prompt model selection makes this easy.
Frequently Asked Questions
Does Ollama really keep everything local?
Yes. Ollama runs models on your hardware and exposes only a local API (localhost:11434). No data is sent to external servers. You can verify this by disconnecting from the internet and confirming the AI features still work.
Which Ollama model is best for grammar checking?
For grammar and spell checking, Deepseek R1 8B offers the best balance of quality and speed. Qwen 3.5 4B is a close second and slightly faster. Deepseek R1 1.5B works for basic grammar but occasionally misses subtle errors.
Can I use Ollama and a cloud provider at the same time?
Yes. WindowSill supports multiple AI providers simultaneously. You can configure Ollama for sensitive tasks and a cloud provider like OpenAI for non-sensitive tasks, then select which model to use on a per-prompt basis.
How much disk space do I need?
A single 7-8B parameter model takes about 4-5 GB of disk space. If you want multiple models available, plan for 10-15 GB. Models are stored in Ollama's data directory and can be removed with ollama rm <model-name>.
Does this work without a GPU?
Yes. Ollama supports CPU-only inference. It's slower (5-15 seconds per response instead of 1-3 seconds with a GPU), but it works. Any modern CPU from the last few years can handle it.