Skip to main content

SQLCLR MCP .NET + Jev Quick Start

Requirements​

  • The .NET 10 SDK to build from source.
  • A reachable SQL Server instance and a login that can read the databases you want to inspect. Use a least-privilege login for real data.
  • An MCP client that can launch a local stdio server. The examples below use Claude Code and Claude Desktop.

Clone or download the SQLCLR MCP .NET + Jev repository. From its root, build and publish the server to a location outside the source checkout:

dotnet build sqlclr-mcp-dotnet.slnx -c Release
dotnet publish sqlclr-mcp-dotnet.csproj -c Release -o "$env:USERPROFILE\.mcp\sqlclr-mcp-dotnet"

Publish the project file explicitly so test projects are not copied into the server directory. On Windows, the launcher is sqlclr-mcp-dotnet.exe.

Connect SQL Server​

Set SQLCLR_CONNECTION_STRING in the environment used to launch the MCP server. For example, with Windows integrated authentication:

Server=localhost;Initial Catalog=AdventureWorks;Integrated Security=True;TrustServerCertificate=True;Encrypt=True

The optional database argument on a tool can select another database on the same server. It does not change the server or credentials.

Register with Claude Code​

Adjust the server, database, and executable path for your computer:

claude mcp add sqlclr -s user -e 'SQLCLR_CONNECTION_STRING=Server=localhost;Initial Catalog=AdventureWorks;Integrated Security=True;TrustServerCertificate=True;Encrypt=True' -- 'C:\Users\<you>\.mcp\sqlclr-mcp-dotnet\sqlclr-mcp-dotnet.exe'
claude mcp list

You can also register the executable in a client's MCP JSON configuration. The repository README includes a Claude Desktop example.

First questions​

Ask your MCP client to list the tables in AdventureWorks, describe a table, then review the schema. To select a specific tool, ask it to call review_database_schema with the database name and an optional schema filter.

Jev is disabled by default. Local catalog review and profiling work without a TypeSafe account. See Architect Analysis before enabling semantic judgments, and Security and Data for what each tool can reveal.