mcp-pwsh

A persistent, pooled PowerShell MCP server. It gives an MCP client (Claude Desktop, or any Model Context Protocol host) a pool of long-lived, named PowerShell 7 sessions with clean stdin → stdout I/O — like keeping several real consoles open, instead of firing one-shot commands.

Why

A naive "run a PowerShell command" tool spawns a fresh process every call, so state does not survive (cwd, variables, imported modules), and the output comes back wrapped in host noise. mcp-pwsh instead keeps one real pwsh process per named session:

• Persistent state — cwd, variables, $env:, and imported modules survive between calls within a session.

• Clean output — a bootstrap loop reads commands with [Console]::In.ReadLine() (no prompt, no echo) and delimits each result with a runtime marker, so you get exactly the command's output plus its exit code — no PS> noise.

• Named session pool (lanes) — each session is an isolated lane with its own process, state, and lock. Different sessions run independently, so two tasks can work at once without corrupting each other's state.

• No escaping games — commands are sent base64-encoded over one line, so quoting, newlines, and multi-line scripts just work.

Related MCP server: SitecoreMCP

Install

pip install mcp-pwsh

Requires Python 3.9+ and PowerShell 7 (pwsh). On Windows the default pwsh location is auto-detected; otherwise pwsh is looked up on PATH.

Configure in an MCP client

Using the installed console script (recommended):

{
  "mcpServers": {
    "pwsh": {
      "command": "mcp-pwsh"
    }
  }
}

Or via the module (useful to pin a specific interpreter):

{
  "mcpServers": {
    "pwsh": {
      "command": "C:\\path\\to\\python.exe",
      "args": ["-m", "mcp_pwsh"]
    }
  }
}

Tools

Discipline: pick one stable session id per chat, reuse it, and pwsh_close it when finished. Use pwsh_list to see lingering sessions and pwsh_kill_all only as a last resort.

Configuration (environment variables)

How it works

Each session launches pwsh running a small bootstrap loop (passed via -EncodedCommand). The loop reads "<id> <base64-command>" lines from stdin, runs the command in a persistent scope, and prints <<<MCPEND:<id>:<exitcode>>>>. Because the full marker is never present in the input, it can never false-match, and because input is read (not typed into an interactive prompt) there is no echo or prompt to strip. The bootstrap also restores a sane PATHEXT/ComSpec in case the host launched the process with a stripped environment, so native executables run normally.

Self-test

Runs a local check of session isolation, listing, close, and kill-all (needs pwsh available; no MCP host required):

python -m mcp_pwsh --selftest

License

GPL-2.0