Configure an MCP client
The examples below use VS Code on Windows. VS Code stores workspace MCP configuration in .vscode/mcp.json; its current MCP schema uses a top-level servers object and a stdio entry with type, command, and optional env fields. See the VS Code MCP configuration reference.
1. Create a fresh data root
Section titled “1. Create a fresh data root”Before VS Code starts the server, create a new directory and keep its resolved path:
$home = Join-Path $env:TEMP ("memscope-mcp-first-session-" + [guid]::NewGuid())New-Item -ItemType Directory -Force $home | Out-Null$homeA fresh root contains no activated plugin files. The server will create its session-log directory after startup.
2. Add the VS Code workspace configuration
Section titled “2. Add the VS Code workspace configuration”Open a disposable or otherwise selected VS Code workspace. Run MCP: Open Workspace Folder MCP Configuration from the Command Palette, then write .vscode/mcp.json with the absolute path printed above:
{ "servers": { "memscope": { "type": "stdio", "command": "memscope-mcp", "env": { "MEMSCOPE_HOME": "C:\\Users\\you\\AppData\\Local\\Temp\\memscope-mcp-first-session-<new-id>" } } }}command must resolve in the environment where VS Code runs; use the full path to memscope-mcp.exe if it is not on PATH. Keep the generated MEMSCOPE_HOME entry in place before the first start. Activated .py files under that root’s plugins directory would load at server startup.
3. Start and list the server
Section titled “3. Start and list the server”Run MCP: List Servers, select memscope, and start it. Confirm that the client lists exactly 11 tools before selecting a target. The bare command and memscope-mcp server start the same local stdio server.
Other MCP clients
Section titled “Other MCP clients”Other MCP clients use the same executable and environment but may use a different configuration filename and top-level schema. Their equivalent server entry must launch memscope-mcp, pass a fresh MEMSCOPE_HOME before startup, reserve stdin/stdout for MCP traffic, and read diagnostics from stderr. Check the chosen client’s current documentation for its exact configuration location and schema.
Continue with Connect to a process and read its PE signature.