Msvc.Info

VS MCP Server — Project Overview & Development Estimates

What It Is

A Visual Studio 2022 Extension that runs an MCP (Model Context Protocol) server inside Visual Studio, giving AI coding assistants real-time access to Roslyn's semantic understanding of the codebase — the same engine that powers IntelliSense.

Instead of AI tools blindly grepping files, they get structured, semantic access to solution structure, symbol search, inheritance trees, refactoring, build/test execution, and editor context — all through a standardized protocol with 70–90% token savings via compact output formatting.

Architecture

graph TB
    subgraph VS["Visual Studio 2022"]
        EXT["Extension Entry\nVSIX Package"]
        HOST["MCP Server Host\nstart · stop · restart"]
        SDK["MCP SDK Server\nModelContextProtocol"]
        SETTINGS["Settings UI\nToolWindow + XAML"]
        STORE[("JSON Settings\nper-solution .mcp/")]

        EXT --> HOST
        EXT --> SETTINGS
        HOST --> SDK
        SETTINGS --> STORE
    end

    subgraph TOOLS["20 MCP Tools"]
        NAV["Navigation\nSolutionTree · Outline\nProjectRefs"]
        SYM["Symbol Search\nFind · Definition · Usages\nInheritance · Callers"]
        EDIT["Editor Context\nActiveFile · Selection"]
        CMD["Commands\nBuild · Clean · AsyncTests"]
        PATH["Path Translation\nWindows · WSL · URI"]
        REF["Refactoring\nRename · Format"]
    end

    SDK --> TOOLS
    TOOLS -.->|"Roslyn API"| ROSLYN[("Roslyn\nWorkspace")]
    TOOLS -.->|"DTE API"| DTE[("VS DTE\nServices")]

    subgraph TRANSPORT["Transports"]
        HTTP["HTTP\nHttpListener\nport 3001"]
        SSE["SSE\nServer-Sent Events"]
        STDIO["Stdio Bridge\n.NET 8 Console"]
    end

    HOST --> HTTP
    HOST --> SSE
    STDIO -->|"HTTP"| HTTP

    CC["AI Coding CLI"] -->|"HTTP"| HTTP
    GEM["Gemini CLI / Codex"] -->|"HTTP"| HTTP
    CD["AI Desktop App"] -->|"stdio"| STDIO
    OTHER["Aider · Cline\nWindsurf · OpenCode"] -->|"HTTP/SSE"| HTTP
  

Solution Structure

Development Time Estimate

How long would a super senior .NET developer take to build this from scratch, writing all code by hand?

Phase Breakdown

Where Time Actually Goes

Risk Factors

What AI Assistance Changes

Key Observations

1. Research is the hidden time sink. VS Extensibility new model is poorly documented. MCP protocol has edge cases. Roslyn's API surface is enormous. A super senior can navigate faster, but still needs to read and experiment. AI lookup (ChatGPT) saves ~1–2 days vs pure documentation reading.
2. Symbol Search (Roslyn) is the hardest implementation phase. 7 tools all hitting different Roslyn APIs (SymbolFinder, syntax trees, semantic model). Call hierarchy and inheritance require deep understanding of Roslyn's compilation model. 3 days optimistic — this could easily become 5 with edge cases.
3. Custom transport is an unexpected cost. The official MCP SDK doesn't ship an HTTP transport for .NET 4.7.2. Building HttpListener-based transport with proper lifecycle management (restart, port reuse, dispose) is a full 2-day effort that wouldn't be obvious from the initial spec reading.
4. VS debugging is inherently slow. Every test cycle requires launching a VS experimental instance (~30s startup). This multiplies debugging time by ~3x compared to normal .NET development. No amount of skill eliminates this overhead.
5. The settings UI has disproportionate complexity. RemoteUserControl (VS Extensibility's XAML) has no standard converters, limited DataTemplate support, and XAML must be EmbeddedResource with PageRemove. A simple settings dialog takes 2 days instead of the expected 0.5 days.

Actual Development Timeline (with AI Coding Assistant)

The project was built with an AI coding assistant writing ~90% of the code, directed by a senior developer who designed architecture and made all decisions. Human role: architecture, vision, decisions, prompting, review, manual testing in VS. Development happened in sporadic bursts over 9 months, not continuous full-time work.

Development Sprints

Estimate vs Reality

Key Observations

1. Sporadic development works surprisingly well with AI. 9 months elapsed, but only ~10 full days of actual work. The AI assistant's memory bank system meant zero ramp-up time between sessions — even after 3-month gaps, context was restored instantly.
2. Settings UI was the one area with no speedup (1x). VS Extensibility's RemoteUserControl has no standard converters, undocumented XAML quirks, and requires EmbeddedResource + PageRemove workarounds. AI couldn't shortcut the pain because the documentation barely exists. This matches the LSAI experience where LSP debugging resisted AI acceleration.
3. Jan 21 was the monster day: 21 commits. Settings expansion, tool filtering, RenameSymbol, FormatDocument, and VS Marketplace publishing — all in a single day. This kind of throughput is only possible when AI handles the boilerplate while the human focuses on architecture decisions.
4. The real bottleneck is manual VS testing. Every change requires launching a VS experimental instance (~30s), connecting an MCP client, and testing tool responses manually. Unit tests cover logic, but integration testing is inherently slow. This is why the speedup is 2.1x instead of the 5–6x seen in LSAI.
5. Marketing took as much time as some features. 2 days for marketing materials (Dev.to article, LinkedIn, landing page, Twitter) is comparable to building the entire settings system. Non-code work doesn't benefit as much from AI coding assistance.