VS-MCP: Roslyn-Powered MCP Server for Visual Studio

Other tools read your files. VS-MCP understands your code.

The first and only Visual Studio extension that gives AI assistants access to the C# compiler (Roslyn) and the Visual Studio Debugger through the Model Context Protocol. 41 tools. 13 powered by Roslyn. 19 debugging tools (Preview). Semantic understanding, not text matching.

See VS-MCP in Action

Click to watch on YouTube

Why VS-MCP?

AI coding tools (Claude Code, Codex CLI, Gemini CLI, OpenCode, Cursor, Copilot, Windsurf...) operate at the filesystem level — they read text, run grep, execute builds. They don't understand your code the way Visual Studio does.

VS-MCP bridges this gap by exposing IntelliSense-level intelligence and the Visual Studio Debugger as MCP tools. Your AI assistant gets the same semantic understanding that powers F12 (Go to Definition), Shift+F12 (Find All References), safe refactoring — and now runtime debugging.

What becomes possible:

22 Stable Tools

Semantic Navigation (Roslyn-powered)

Code Understanding (Roslyn-powered)

Code Analysis (Roslyn-powered)

Refactoring (Roslyn-powered)

Project & Build

Editor Integration

19 Debugging Tools (Preview)

Your AI assistant can now debug your .NET code at runtime through the Visual Studio Debugger. Set breakpoints, step through code, inspect variables, attach to Docker containers and WSL processes.

Debug Control (10 tools)

Debug Inspection (5 tools)

Breakpoint Management (4 tools)

AI Debugging Guide

Complete reference for AI agents — all 19 tools, 10 workflows, Docker & WSL setup, polling patterns, and best practices.

Download AI Debugging Guide (.md) — add it to your AI's context for full debugging capabilities.

Compatible Clients

Works with any MCP-compatible AI tool:

CLI Agents:

• Claude Code — Anthropic's terminal AI coding agent
• Codex CLI — OpenAI's terminal coding agent
• Gemini CLI — Google's open-source terminal agent
• OpenCode — Open-source AI coding agent (45k+ GitHub stars)
• Goose — Block's open-source AI agent
• Aider — AI pair programming in terminal

Desktop & IDE:

• Claude Desktop — Anthropic's desktop app
• Cursor — AI-first code editor
• Windsurf — Codeium's AI IDE
• VS Code + Copilot — GitHub Copilot with MCP
• Cline — VS Code extension
• Continue — Open-source AI assistant

Any MCP client — open protocol, universal compatibility

Quick Start

1. Install this extension from Visual Studio Marketplace
2. Open your .NET solution in Visual Studio
3. Configure port in MCP Server Settings (default: 3010)
4. Add to your MCP client:

"vs-mcp": {
  "type": "http",
  "url": "http://localhost:3010/sdk/"
}

5. Start asking your AI semantic questions about your code

Works With Any MCP Client

Any tool supporting Model Context Protocol will work.

Configure AI Preferences (Recommended)

Add these instructions to your project's AI config file (see table above) to ensure your AI automatically prefers MCP tools:

## MCP Tools - ALWAYS PREFER

When `mcp__vs-mcp__*` tools are available, ALWAYS use them instead of Grep/Glob/LS:

| Instead of | Use |
|------------|-----|
| `Grep` for symbols | `FindSymbols`, `FindSymbolUsages` |
| `LS` to explore projects | `GetSolutionTree` |
| Reading files to find code | `FindSymbolDefinition` then `Read` |
| Searching for method calls | `GetMethodCallers`, `GetMethodCalls` |

**Why?** MCP tools use Roslyn semantic analysis - 10x faster, 90% fewer tokens.

This ensures your AI uses the semantic tools without needing to say "use vs-mcp" in every prompt.

Multi-Solution: Understand Your Dependencies

Need the AI to understand a library you depend on? Clone the source from GitHub, open it in a second Visual Studio — each instance runs its own MCP server on a configurable port.

Your project                          → port 3010
Library source (cloned from GitHub)   → port 3011
Framework source                      → port 3012

Your AI connects to all of them. It can trace calls, find usage patterns, and understand inheritance across your code and library code. No more guessing how a dependency works — the AI reads its source with compiler-level understanding.

MCP client configuration — three options depending on your client:

Clients with native HTTP support (Claude Desktop, Claude Code):

{
  "mcpServers": {
    "vs-mcp": {
      "type": "http",
      "url": "http://localhost:3010/sdk/"
    },
    "vs-mcp-whisper": {
      "type": "http",
      "url": "http://localhost:3011/sdk/"
    }
  }
}

Clients without HTTP support (via mcp-remote proxy):

{
  "mcpServers": {
    "vs-mcp": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "http://localhost:3010/sdk/"]
    },
    "vs-mcp-whisper": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "http://localhost:3011/sdk/"]
    }
  }
}

Codex CLI (TOML):

[mcp_servers.vs-mcp]
type = "stdio"
command = "npx"
args = ["mcp-remote", "http://localhost:3010/sdk/"]

[mcp_servers.vs-mcp-whisper]
type = "stdio"
command = "npx"
args = ["mcp-remote", "http://localhost:3011/sdk/"]

Each name tells the AI what codebase it's connecting to. Simple, explicit, powerful.

Settings

Tools → MCP Server Settings

Changes apply on next server start.

Example Prompts

Once connected, just ask naturally. Add (vs-mcp) to hint which server to use — after the first time, AI gets it:

Find all usages of UserService (vs-mcp)

Rename ProcessOrder to HandleOrderAsync across the solution

Show me the inheritance hierarchy of BaseController

What methods call ValidateUser?

Run the tests in PaymentTests and tell me what's failing

Set a breakpoint on OrderService.ProcessOrder and start debugging

If you have multiple MCP servers connected (e.g., vs-mcp + vs-mcp-whisper), specify which one when needed.

Who Is This For?

Enterprise .NET developers working on solutions with:

• Multiple projects and complex architectures
• Inheritance hierarchies and interface implementations
• Dependency injection (where grep can't find implementations)
• Large codebases where text search produces too much noise
• Runtime bugs that require actual debugging, not guessing

If your AI assistant keeps saying "I found 47 matches for that method name" — you need VS-MCP.

vs. Other Approaches

Requirements

• Visual Studio 2022 (17.13+) or Visual Studio 2026
• Windows (amd64 or arm64)
• Any MCP-compatible AI tool

Support

• Issues & Feature Requests: GitHub Issues
• Questions: Marketplace Q&A
• MCP Protocol: modelcontextprotocol.io

About

0ics srl — Italian software company specializing in AI-powered development tools. Part of the example4.ai ecosystem.

Built by Ladislav Sopko — 30 years of software development, from assembler to enterprise .NET.

VS-MCP: Because your AI deserves the same intelligence as your IDE.