English · 한국어

FieldCure AssistStudio

AI Chat Controls for WinUI 3 + Windows AI Workspace App — Bring Your Own Key, pick any provider.

Adapter, runner, and MCP-server packages have their own badges in Packages.

AssistStudio is two things:

1. A reusable WinUI 3 library for building desktop AI assistants with multi-provider support, tool approval, and profile-based behavior. Comes with optional adapter packages so you can stream from a vendor SDK (e.g. the official Anthropic SDK) directly into the chat UI when you want its latest features without reimplementing IAiProvider.
2. A Windows-native AI workspace app for cloud and local models, profiles, tools, and structured conversations.

Contents

• Features
• Architecture
• Quick Start
• Providers
• Controls
• MCP Integration
• Configuration
• Requirements
• Packages

Architecture Overview

---

Features

Providers & Streaming

• BYOK (Bring Your Own Key) — Users supply their own API keys. No proxy, no middleman.
• Multi-Provider — Claude, OpenAI, Gemini, Ollama, Groq, and any OpenAI-compatible endpoint out of the box. Implement IAiProvider to add your own.
• Streaming — Real-time structured event streaming via IAsyncEnumerable<StreamEvent> with elapsed time display.
• Extended Thinking — Per-provider thinking/reasoning support (Claude, OpenAI o-series, Ollama native thinking). Configurable via ThinkingOverride and ThinkingBudget.
• Token Tracking — Input/output token counts exposed after every request.

Tools & Agents

• Tool / Function Calling — Define tools with IAssistTool. ToolCallExecutor orchestrates execution with confirmation flow and parallel execution.
• MCP Integration — Connect to MCP servers (Stdio / HTTP) to aggregate tools. McpToolAdapter bridges MCP tools to IAssistTool.
• Built-in MCP Servers — Essentials (12–16 tools), Filesystem, Knowledge Base (RAG), Outbox, and Runner — fetched and run directly from NuGet via dnx (.NET 10 SDK), pinned to a major-version range so minor/patch updates flow through on the next launch.
• Sub-Agent Delegation — delegate_task for autonomous sub-agent execution. Parallel dispatch (multiple delegate_task blocks in a single model turn run concurrently and each renders a pulsing placeholder that resolves in place), specialist routing (ISpecialist — Web Search, Judgment), and truncation-aware result handling (sub-agent max_tokens cutoff is propagated as status: "truncated" so the parent never forwards a mid-markdown stub as final).

UI & Media

• Re-templatable WinUI 3 Controls — ChatPanel, ComposeBar, AttachmentPreviewBar, ToolApprovalPanel — all TemplatedControls with Generic.xaml override.
• Conversation Branching — Tree-based message editing with branch navigator (◀ 1/2 ▶). Edit any message to explore alternatives without losing history.
• Vision & Documents — Attach images, PDFs, DOCX, HWPX, XLSX, PPTX. Automatic image compression. Per-provider PdfCapability.
• Multimedia Tool Results — MCP image, audio, and video content rendered inline. Image hover toolbar with zoom, save, and copy.
• Localization — Built-in en-US and ko-KR resource strings.

Data & Persistence

• Conversation Persistence — Save and load conversations in .astx (ZIP archive) format with full branching tree and media persistence.
• Profiles & Presets — Save provider configurations as presets; switch system prompts and tool selections with profiles.
• Knowledge Base — Multi-KB management with embedding model selection and per-conversation KB selector.
• Schedule — Cron schedule management with bilingual (en/ko) human-readable descriptions.
• Workspace Context — IWorkspaceContext for dynamic system prompt injection based on app state.
• Structured Logging — DiagnosticLogger with pluggable OnException, OnWarning, OnInfo callbacks.

---

Screenshots

Empty State	Tool Approval

---

Architecture

Project	NuGet Package	TFM	Key Types
Ai.Providers	FieldCure.Ai.Providers	net8.0	IAiProvider, StreamEvent, IAssistTool, AiRequest, AiResponse, ChatMessage, ProviderModel, CustomProviderConfig, McpToolAdapter, ImageCompressor, IMultiContentTool, MediaContent
Ai.Execution	FieldCure.Ai.Execution	net8.0	IAgentLoop, AgentLoop, ISubAgentExecutor, SubAgentExecutor, AgentLoopContext, AgentLoopResult
AssistStudio.Core	FieldCure.AssistStudio.Core	net8.0	ISpecialist, KnowledgeBase, ToolCallExecutor, ToolResolver, IWorkspaceContext, BuiltInServerConfig, Profile
AssistStudio.Controls	FieldCure.AssistStudio.Controls.WinUI	net8.0-windows10.0.19041.0 net10.0-windows10.0.19041.0	ChatPanel, ComposeBar, AttachmentPreviewBar, ToolApprovalPanel, ToolElicitationPanel, ChatTheme
AssistStudio.Anthropic (optional adapter)	FieldCure.AssistStudio.Anthropic	net8.0 net10.0	AnthropicStreamEventMapper, AnthropicMessageConverter — maps the official Anthropic SDK's RawMessageStreamEvent to StreamEvent
AssistStudio.Controls.Anthropic (optional adapter)	FieldCure.AssistStudio.Controls.WinUI.Anthropic	net8.0-windows10.0.19041.0 net10.0-windows10.0.19041.0	ChatPanelExtensions — one-call streaming of an Anthropic SDK response into a ChatPanel via AssistantTurnHandle
AssistStudio	(workspace app)	net10.0-windows10.0.19041.0	Reference implementation with settings, MCP server management, sub-agent delegation, schedule, and PasswordVaultHelper

Core is platform-agnostic (net8.0). It has no Windows-specific dependencies — you can reference it from a console app, a server, or any .NET project.

See Dependency Graph for the full cross-repository package dependency map.

Choose your path:

• Building with the libraries — start with Quick Start below; the Controls package README covers ChatPanel template parts and the full property reference.
• Running the workspace app — install from the Microsoft Store, or clone this repo and dotnet run --project src/AssistStudio.

---

Quick Start

1. Install packages

dotnet add package FieldCure.AssistStudio.Core
dotnet add package FieldCure.AssistStudio.Controls.WinUI

2. Create a provider and wire up the control

using FieldCure.Ai.Providers;

// Pick a provider — API key comes from the user
var provider = new ClaudeProvider(apiKey: "sk-ant-...", modelId: "claude-sonnet-4-6");

<!-- In your WinUI 3 Page -->
<Page xmlns:assist="using:FieldCure.AssistStudio.Controls">

    <assist:ChatPanel x:Name="Chat"
                      Placeholder="Ask anything..."
                      Theme="System" />
</Page>

// Code-behind
Chat.Provider = provider;

That's it — you have a fully functional AI chat with streaming, Markdown rendering, syntax highlighting, thinking blocks, and conversation branching.

3. Streaming with StreamEvent

var request = new AiRequest("Explain quantum computing.");

await foreach (var evt in provider.StreamAsync(request))
{
    switch (evt)
    {
        case StreamEvent.ThinkingDelta t:
            Console.Write($"[think] {t.Text}");
            break;
        case StreamEvent.TextDelta d:
            Console.Write(d.Text);
            break;
        case StreamEvent.ToolCallStart s:
            Console.WriteLine($"\n→ Calling {s.FunctionName}...");
            break;
        case StreamEvent.Usage u:
            Console.WriteLine($"\nTokens: {u.TokenUsage.TotalTokens}");
            break;
    }
}

---

Providers

Supported providers

Provider	Streaming	Vision	Documents	Tool Calling	Thinking	API Key Required
Claude (Anthropic)	Yes	Yes	Yes	Yes	Yes	Yes
OpenAI (+ compatible)	Yes	Yes	Yes	Yes	o-series	Yes
Gemini (Google)	Yes	Yes	Yes	Yes	No	Yes
Ollama (local)	Yes	Dep.	Dep.	Dep.	think tags	No
Groq	Yes	Yes	Yes	Yes	Dep.	Yes
Custom (OpenAI-compatible)	Yes	Dep.	Dep.	Dep.	think tags	Dep.

Register any OpenAI-compatible endpoint via ProviderFactory.RegisterCustomProvider with a CustomProviderConfig (BaseUrl, DisplayName).

Implementing a custom provider

Implement IAiProvider to integrate any AI service. The interface boils down to:

public interface IAiProvider
{
    string ProviderName { get; }
    string ModelId { get; }
    PdfCapability PdfCapability { get; }
    TokenUsage? LastUsage { get; }

    Task<AiResponse> CompleteAsync(AiRequest request, CancellationToken ct = default);
    IAsyncEnumerable<StreamEvent> StreamAsync(AiRequest request, CancellationToken ct = default);
    Task<IReadOnlyList<AiModel>> ListModelsAsync(CancellationToken ct = default);
    Task<ConnectionInfo> ValidateConnectionAsync(CancellationToken ct = default);
    ThinkingSupport GetThinkingSupport(string modelId);
}

Assign your implementation to a ChatPanel:

Chat.Provider = new MyCustomProvider();

See docs/CustomProvider.md for a complete walkthrough — full skeleton, streaming, tool calling, and how to register it with ProviderFactory.

Third-party SDK adapters

When you'd rather stream from an official vendor SDK instead of using a built-in IAiProvider implementation, the adapter packages bridge that SDK's event stream into ChatPanel without you having to write an IAiProvider.

Anthropic SDK — FieldCure.AssistStudio.Anthropic (platform-agnostic, for any .NET host) and FieldCure.AssistStudio.Controls.WinUI.Anthropic (ChatPanel extensions):

using Anthropic;
using FieldCure.AssistStudio.Controls.Anthropic;

var client = new AnthropicClient();
var sdkStream = client.Messages.CreateStreaming(new MessageCreateParams
{
    Model = "claude-sonnet-4-6",
    MaxTokens = 4096,
    Messages = [/* … */],
});

// Begin an assistant turn, then consume the SDK stream through the mapper.
// RawMessageStreamEvent is translated to StreamEvent and rendered directly.
using var turn = Chat.BeginAnthropicTurn(providerName: "Claude", modelId: "claude-sonnet-4-6");
await turn.StreamAnthropicAsync(sdkStream);

Use the adapter when you need features the vendor SDK ships first (new beta capabilities, prompt caching controls, etc.) while keeping the ChatPanel UX — branching, attachments, tool approval — intact.

---

Controls

All controls are TemplatedControls defined in Generic.xaml. They carry no app-level dependency — reference the NuGet package and use them in any WinUI 3 project.

ChatPanel

The main control. Provides a complete chat experience: message list (WebView2), input area, streaming, thinking blocks, conversation branching, attachments, presets, and profiles.

<assist:ChatPanel Provider="{x:Bind ViewModel.Provider, Mode=OneWay}"
                  SystemPrompt="You are a helpful assistant."
                  Theme="Dark"
                  Placeholder="Type a message..."
                  AvailableModels="{x:Bind ViewModel.Models}"
                  SelectedModel="{x:Bind ViewModel.CurrentModel, Mode=TwoWay}"
                  RegisteredTools="{x:Bind ViewModel.Tools}"
                  WorkspaceContext="{x:Bind ViewModel.Workspace}" />

See the Controls package README for the full dependency-property reference (provider/model, workspace, knowledge base, tools, behavior, theming) and the events list. The same content is published on the NuGet package page.

ComposeBar

The chat input area — text box, attach button, model/profile selectors. Used internally by ChatPanel, but can be placed standalone.

AttachmentPreviewBar

Horizontal scrollable bar showing thumbnails of attached files before sending.

ToolApprovalPanel

Inline confirmation panel shown when a tool with RequiresConfirmation = true is invoked. Displays the tool name, an expandable JSON arguments preview, and Allow/Reject buttons. Replaces ComposeBar during confirmation and restores it after.

ToolElicitationPanel

Inline input panel shown when an MCP server issues an elicitation/create request mid-tool-call (e.g. asking for a missing API key or confirming a destructive action). Renders the server-supplied JSON Schema as a native form — string, number, boolean, enum, and multi-field objects. String fields whose name looks secret (password, secret, token, apiKey, etc.) automatically render as PasswordBox to avoid leaking values via shoulder-surfing or screen capture. Server-origin badge and batch submit keep the flow auditable.

Re-templating

Override the default template in your app's resources:

<Style TargetType="assist:ChatPanel" BasedOn="{StaticResource DefaultChatPanelStyle}">
    <!-- Customize template parts: PART_MessageList, PART_InputArea, PART_TitleBar, etc. -->
</Style>

---

MCP Integration

The workspace app demonstrates full MCP (Model Context Protocol) integration:

• McpServerRegistry — Singleton server connection manager with observable tool collection
• McpServerConnection — Per-server lifecycle (Disconnected → Connecting → Connected / Error)
• McpToolAdapter — Bridges MCP tools to IAssistTool without Core SDK dependency
• ToolResolver — Merges built-in + MCP tools, prefixing MCP tool names with server name on conflict
• SearchToolsTool — Meta-tool for searching across large MCP tool sets to save tokens

MCP servers are configured with Stdio or HTTP transport and connected at app startup. Tools from all connected servers are aggregated and made available to providers.

Built-in MCP Servers

The workspace app bundles a set of MCP servers (Essentials, Filesystem, Knowledge Base, Outbox, Runner) — see Packages → MCP Servers below for the full list with descriptions and badges.

BuiltInServerHelper launches each built-in server as dnx <PackageId>@<Major>.* --yes … at startup. dnx fetches and caches the latest in-range version directly from NuGet, so a new in-range build is picked up automatically on the next server spawn. Read-only tools (read, list, search) skip user approval.

Update model

No manual dotnet tool update step is needed. dnx resolves within the pinned major range on each invocation, so minor/patch releases flow through automatically on the next app launch. Major-version bumps (e.g. 2.* → 3.*) are intentionally release-coupled to AssistStudio itself so breaking changes can be validated against the host before being picked up.

---

Configuration

Provider models

var model = new ProviderModel
{
    Name = "Claude Sonnet",
    ProviderType = "Claude",
    ApiKey = "sk-ant-...",
    ModelId = "claude-sonnet-4-6",
    Temperature = 0.7,
    MaxTokens = 4096,
    StreamingEnabled = true,
    ThinkingEnabled = true,
    ThinkingBudget = 8192
};

Profiles

Profiles pair a system prompt with optional provider/model preferences and tool selections:

var profile = new Profile
{
    Name = "Task Planner",
    SystemPrompt = "You are a task planner that breaks down complex requests into steps.",
    EnabledServers = ["builtin_essentials", "builtin_filesystem"],
    UseSearchTools = true  // Use meta-tool instead of sending all tool definitions
};

---

Requirements

Dependency	Minimum Version	Scope
.NET runtime	8.0	Library consumers (Core / Controls / Providers / Execution)
.NET SDK	10.0	AssistStudio app — dnx (ships with the .NET 10 SDK) launches built-in MCP servers
Windows App SDK	1.7	AssistStudio app + Controls
WebView2 Runtime	Evergreen	AssistStudio app + Controls
Target Platform	Windows 10 1903+ (10.0.19041.0)	AssistStudio app + Controls

---

Packages

MCP Servers

Package	Version	Description
FieldCure.Mcp.Essentials		12–16 tools — HTTP, web search (+ news/images/scholar/patents), web/document fetching, shell, JavaScript sandbox, environment info, file I/O, persistent memory
FieldCure.Mcp.Outbox		Multi-channel messaging — Slack, Discord, Telegram, Email (Gmail / Naver / SMTP / Microsoft Graph), KakaoTalk
FieldCure.Mcp.Filesystem		Sandboxed file/directory operations with built-in document parsing (DOCX, HWPX, XLSX, PDF)
FieldCure.Mcp.Rag		Document search — hybrid BM25 + vector retrieval, multi-KB, incremental indexing
FieldCure.Mcp.PublicData.Kr		Korean public data gateway — data.go.kr (80,000+ APIs)
FieldCure.AssistStudio.Runner		Headless LLM task runner with scheduling via Windows Task Scheduler. Windows-only.

Libraries

Package	Version	Description
FieldCure.Ai.Providers		Multi-provider AI client — Claude, OpenAI, Gemini, Ollama, Groq with streaming and tool use
FieldCure.Ai.Execution		Agent loop and sub-agent execution engine for autonomous tool-use workflows
FieldCure.AssistStudio.Core		MCP server management, tool orchestration, and conversation persistence
FieldCure.AssistStudio.Controls.WinUI		WinUI 3 chat UI controls — WebView2 rendering, streaming, conversation branching
FieldCure.AssistStudio.Anthropic		Platform-agnostic Anthropic SDK adapter — maps RawMessageStreamEvent to StreamEvent, converts ChatMessage ↔ MessageParam
FieldCure.AssistStudio.Controls.WinUI.Anthropic		WinUI 3 ChatPanel extensions for streaming Anthropic SDK responses via AssistantTurnHandle
FieldCure.DocumentParsers		Document text extraction — DOCX, HWPX, XLSX, PPTX with math-to-LaTeX, plus PdfPig-backed PDF text (auto-registered from v2.0)
FieldCure.DocumentParsers.Imaging		PDF page-as-image rendering via PDFium. Needed by hosts that send PDFs to vision-capable providers (register with AddImagingSupport())
FieldCure.DocumentParsers.Ocr		Tesseract-based OCR for scanned PDFs and image documents. Windows x64 only

---

Contributing

Contributions are welcome! Please open an issue first to discuss what you'd like to change.

1. Fork the repository
2. Create a feature branch (git checkout -b feature/amazing-feature)
3. Commit your changes
4. Push to the branch and open a Pull Request

---

License

MIT — Copyright (c) 2026 FieldCure Co., Ltd.