Sample: Foundry Local + .NET MAUI on-device AI

[davidortinau]

Summary

Guide and code samples showing how to integrate Foundry Local (on-device AI) into a .NET MAUI app alongside Microsoft Foundry (cloud), with a runtime toggle between providers.

What's included

• Architecture overview with IChatClient provider swap pattern
• Platform-specific integration:
  • Windows: Native SDK via Microsoft.AI.Foundry.Local.WinML
  • Mac Catalyst: REST API discovery via Foundry Local CLI (no NuGet needed)
  • iOS: Notes on using ONNX Runtime GenAI directly
• ThinkingTagStripperChatClient middleware for models that emit <think> reasoning tags
• Manual JSON extraction for reliable structured output from small local models
• Settings UI toggle with download progress
• Offline detection and auto-fallback patterns
• Model selection guide and platform support matrix

Tested end-to-end on

• macOS (Mac Catalyst) with qwen3-0.6b via Foundry Local CLI
• Streaming chat completions, structured JSON output, tool calling

Related

• Blog draft: "On-Device AI in .NET MAUI with Foundry Local" (forthcoming on devblogs)
• Foundry Local repo: https://github.com/microsoft/Foundry-Local

/cc @mattleibow @jfversluis

[jfversluis]

Code Multi-Model Consensus (Sonnet 4.5, GPT-5.2, Opus 4.6)Review

**1. Mac Catalyst won't no NuGet package for FoundryLocalManager**compile
Code uses FoundryLocalManager under #if WINDOWS || MACCATALYST (lines ~176-197), but the csproj only adds Microsoft.AI.Foundry.Local.WinML conditioned on -windows. The guide says macOS needs "no NuGet package" yet the C# directly references types from Microsoft.AI.Foundry.Local. The Mac build will fail on unresolved types.

Fix: Either add a Mac Catalyst package reference, or split the code paths so MACCATALYST uses a CLI-based discovery approach instead of the managed API.

---

2. AddChatClient(async sp => ...) won't compile
The alternative DI registration example (line ~584) uses an async factory lambda. Standard .NET DI factory delegates are synchronous (Func<IServiceProvider, there is no Func<IServiceProvider, Task> overload.T>)

Fix: Remove this example (the recommended IChatClientService pattern already handles async init correctly), or show a sync factory with .GetAwaiter().GetResult().

---

3. Missing OpenAI NuGet package in csproj snippet
Code uses using OpenAI; and new OpenAIClient(...), but the csproj snippet (lines ~65-68) only lists Azure.AI.OpenAI, Microsoft.Extensions.AI, and Microsoft.Extensions.AI.OpenAI. Readers following the guide will get compile errors.

Fix: Add the OpenAI package to the csproj snippet.

---

4. Mismatched package versions
Microsoft.Extensions.AI is shown at 9.9.1 while Microsoft.Extensions.AI.OpenAI is at 9.4.4-preview. These packages should be version-aligned to avoid binding issues.

---

5. ProgressBar value range mismatch
DownloadProgress = p.PercentComplete; (line ~676) assigns a 100 value, but MAUI ProgressBar.Progress expects 0.1.0. The bar will appear full as soon as download reaches 1%.00

Fix: DownloadProgress = p.PercentComplete / 100.0;

---

6. JSON extraction regex fails on nested objects
The generic GetStructuredResponseAsync<T> helper (line ~951) uses (\{[\s\S]*?\}) (lazy quantifier), which matches from the first { to the nearest }. For any JSON with nested objects this produces invalid JSON.

Fix: Use a greedy match (\{[\s\S]*\}) or implement brace-balancing extraction.

---

**7. async void crash risk + lazy resolution**ConnectivityWatcher
The OnConnectivityChanged handler is async void with no try/catch. Unhandled exceptions will crash the app. Additionally, the singleton is registered but never explicitly resolved, so the constructor (which hooks the event) may never run.

Fix: Add try/catch in the handler, and ensure the watcher is resolved at startup.

---

Review generated with Claude Sonnet 4.5, GPT-5.2, and Claude Opus 4.6.