Microsoft now ships local AI models you can run on your own hardware. No API keys, no cloud accounts, no data leaving your PC. The project is called AI Foundry Local, and it's designed to work with the Windows ecosystem from the ground up.
I added Foundry Local support to WindowSill because a lot of users work in Microsoft-heavy environments. They run Windows, use Office, deploy through Azure, and their IT departments have opinions about where data goes. For those users, a Microsoft-backed local AI option just makes sense. It fits the stack they already trust.
This guide walks you through setting up Foundry Local as a private AI writing assistant in WindowSill. You'll get grammar checking, text rewriting, tone adjustment, and translation running on your machine, powered by Microsoft's own models.
What is Microsoft AI Foundry Local?
AI Foundry Local is Microsoft's open-source tool for running small language models (SLMs) directly on Windows. It manages model downloads, handles hardware optimization (CPU, GPU, and NPU), and exposes a local OpenAI-compatible API that apps like WindowSill can connect to.
The key difference from other local AI tools: Foundry Local is built by Microsoft, ships models from the Microsoft ecosystem (Phi-4, Phi-3.5, and others), and integrates with Windows features like DirectML for hardware acceleration. If your organization already standardizes on Microsoft tooling, this is the local AI option that fits without introducing another vendor.
What you'll need
- WindowSill installed from the Microsoft Store (AI features require WindowSill+)
- Windows 10 or later (Windows 11 recommended for NPU support)
- 8 GB of RAM minimum (16 GB recommended)
- A few gigabytes of disk space for the model you choose
Foundry Local works on CPU, GPU (NVIDIA and AMD via DirectML), and NPU (on supported hardware like Copilot+ PCs). No dedicated GPU is required, but responses will be faster with one.
Step 1: Install Foundry Local
The fastest way to install is through winget. Open a terminal and run:
winget install Microsoft.AI.Foundry.Local
You can also download it from foundrylocal.ai.
Once installed, verify it's working:
foundry --version
If you see a version number, you're set.
Step 2: Download a model for writing tasks
Foundry Local manages models through its CLI. You need to download at least one model before WindowSill can use it.
For writing tasks (grammar, rewriting, tone adjustment), you want a model that handles natural language well. Here are three options from the Foundry Local catalog:
| Model | Size | Best for | Download command |
|---|---|---|---|
| Phi-4-mini | ~2.4 GB | General writing, grammar, rewriting | foundry model download phi-4-mini |
| Phi-3.5-mini | ~2.2 GB | Lighter option, solid grammar | foundry model download phi-3.5-mini |
| Phi-4 | ~8.4 GB | High-quality rewrites, complex tasks | foundry model download phi-4 |
For most people, Phi-4-mini is the right starting point. It's small enough to run comfortably on 8 GB of RAM and handles grammar checks, spell fixes, and short rewrites well.
foundry model download phi-4-mini
This downloads the model to your local cache. It only needs to happen once.
A note on model size vs. quality: Larger models produce better output for complex writing tasks, but they need more RAM and benefit from a GPU. For quick grammar fixes and tone adjustments, Phi-4-mini is more than enough. If you find yourself doing long rewrites or multi-paragraph edits, consider Phi-4 (with a GPU) for noticeably better results.
You can list all available models with:
foundry model list
And see which models you've already downloaded:
foundry model list --cached
Step 3: Connect WindowSill to Foundry Local
WindowSill detects Foundry Local automatically. When you select it as a provider, WindowSill starts the Foundry Local service in the background and loads your chosen model.
In WindowSill:
- Open the Settings from the command bar
- Go to the AI Writing & Analysis section
- Under AI Providers, select Foundry Local
- Pick the model you downloaded (e.g., Phi-4-mini)
That's it. WindowSill handles starting and stopping the Foundry Local service for you. No need to manage it manually in a terminal.
Step 4: Test it
Open any app where you write: Word, Outlook, Teams, a browser, whatever you use. 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 comes 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 Foundry Local loads the model into memory. After that, responses are faster.
What you can do with local AI
Once connected, all of WindowSill's AI writing features work through your local Foundry Local model:
Grammar and spell check. Select text in any app, fix errors without switching windows.
Rewriting. Highlight a paragraph and ask for a rewrite. Good for polishing drafts or simplifying dense writing.
Tone adjustment. Switch between professional, casual, and attention-grabbing tones. Create custom tone presets for recurring needs like "customer support reply" or "executive summary."
Translation. Select text and translate to any of 35+ supported languages. Quality depends on the model. Phi-4-mini handles common language pairs (English/Spanish, English/French, English/German) well. For less common pairs, a larger model helps.
Custom prompts. Build reusable prompts with variable injection. A prompt that takes selected text and converts it into a formatted meeting recap with today's date auto-inserted, for example.
Summarization. Select a long email or document section and get a concise summary.
All of this runs on your machine. Nothing goes to Azure, OpenAI, or any external server.
Why Foundry Local over other local AI tools?
WindowSill also supports Ollama and LM Studio for local AI. So why choose Foundry Local?
Microsoft ecosystem fit. If your organization runs on Windows, Office, and Azure, adding another open-source tool to the stack can mean another thing for IT to approve, audit, and support. Foundry Local is a Microsoft product. That often makes the procurement conversation shorter.
DirectML hardware acceleration. Foundry Local uses DirectML, which means it works with NVIDIA, AMD, and Intel GPUs without needing vendor-specific drivers or CUDA installations. On Copilot+ PCs, it can also use the NPU for inference.
Optimized Microsoft models. Phi-4-mini and Phi-3.5-mini are Microsoft's own small language models, optimized to run well on consumer hardware. Foundry Local ships them in formats tuned for the hardware it detects on your system.
Managed lifecycle. WindowSill starts and stops the Foundry Local service automatically. With Ollama, you manage the background process yourself (or let it run all the time). Foundry Local's service management is handled for you.
That said, Ollama has a larger model library, and LM Studio offers a visual model browser. If model variety is your priority, those are solid choices too. WindowSill supports all three, so you can try each one and see which fits your workflow.
Performance tips
Local models are slower than cloud APIs. Here's how to keep things comfortable:
- Let the model stay loaded. Foundry Local keeps models in memory for faster repeat requests. Avoid restarting the service between tasks if you're actively writing.
- Use Phi-4-mini for quick tasks. Grammar checks and spell fixes don't need a large model. Save the heavier models for complex rewrites.
- Use per-prompt model selection. WindowSill lets you assign different models to different tasks. Use a fast, small local model for grammar and a cloud model for complex rewrites if you want the best of both worlds.
- Check your hardware acceleration. Foundry Local automatically picks the best backend (CPU, GPU, or NPU). If responses feel slow, confirm your GPU drivers are up to date so DirectML can use your hardware properly.
- 16 GB of RAM makes a difference. With 8 GB, the model works but your system might feel sluggish when it's loaded alongside a browser and Office apps. 16 GB gives comfortable headroom.
When local isn't enough
I want to be honest about the trade-offs. Local models are good for most daily writing tasks, but they're not as capable as the latest cloud models for everything.
- 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.
You don't have to pick one. WindowSill supports both local and cloud providers at the same time. Route sensitive content through Foundry Local and non-sensitive content through a cloud API. Per-prompt model selection makes this easy.
Frequently Asked Questions
Does AI Foundry Local keep everything on my machine?
Yes. Foundry Local runs models locally and exposes only a local API endpoint. No data is sent to Microsoft or any external server. You can verify this by disconnecting from the internet and confirming the AI features still work.
What's the difference between Foundry Local and Ollama for WindowSill?
Both run AI models locally. Foundry Local is Microsoft-backed, uses DirectML for hardware acceleration (including NPU on Copilot+ PCs), and ships optimized Microsoft models. Ollama has a larger community model library and a different management CLI. WindowSill supports both, so you can try each and pick what works best for your setup.
Can I use Foundry Local and a cloud provider at the same time?
Yes. WindowSill supports multiple AI providers simultaneously. Configure Foundry Local for sensitive tasks and a cloud provider like OpenAI or Anthropic for non-sensitive tasks, then select which model to use on a per-prompt basis.
How much disk space do I need?
Phi-4-mini takes about 2.4 GB. Larger models like Phi-4 need around 8.4 GB. You can check your downloaded models with foundry model list --cached and remove unused ones to free up space.
Does this work on Copilot+ PCs with an NPU?
Yes. Foundry Local detects available hardware (CPU, GPU, NPU) and picks the best option automatically. On Copilot+ PCs with an NPU, it can offload inference to the NPU for better performance and lower power use.
If you're already running WindowSill and want to keep your writing private, download Foundry Local and try it. The setup takes about five minutes, and you can always switch back to a cloud provider if local doesn't fit your workflow.