A shell MCP server for AI that actually holds a session. Load Import-Module Az once and let AI run 50 follow-up cmdlets in milliseconds each. Watch every command happen in a real terminal window — the same one you can type into yourself.
Prerequisite: .NET 9 Desktop Runtime. No global install needed — npx fetches a ~5 MB native binary on first run.
claude mcp add-json splash -s user '{"command":"npx","args":["-y","splashshell@latest"]}'Claude Desktop
Add to %APPDATA%\Claude\claude_desktop_config.json:
{
"mcpServers": {
"splash": {
"command": "npx",
"args": ["-y", "splashshell@latest"]
}
}
}The @latest tag is important: without it, npx will happily keep reusing a stale cached copy even after a new version ships.
Other shell MCP servers are either stateless (fresh subshell per command, nothing persists) or headless (persistent PTY, but invisible to you). SplashShell is neither — and that unlocks things the others can't do.
Session persistence helps every shell, but for PowerShell it's transformative. Most MCP shell servers spin up a fresh subshell per command — which makes real PowerShell workflows impractical:
- 10,000+ modules on PowerShell Gallery. Az (Azure), AWS.Tools, Microsoft.Graph (Entra ID / M365), ExchangeOnlineManagement, PnP.PowerShell, SqlServer, ActiveDirectory — plus every CLI in PATH (git, docker, kubectl, terraform, gh, az, aws, gcloud) and full access to .NET types.
- 30–70 second cold imports, paid once.
Import-Module Az.Compute, Az.Storage, Az.Networkcan take over a minute on the first call. A subshell-per-command MCP server pays that cost on every command and the AI gives up on Azure workflows entirely. With SplashShell, the AI imports once and every subsequent cmdlet runs in milliseconds. - Live .NET object graphs. PowerShell pipes rich objects, not text. After
$vms = Get-AzVM -Status, the AI can chain arbitrary follow-ups against the live object — filter, group, drill into nested properties — without re-hitting Azure. In a one-shot MCP server, that object vanishes the moment the command returns. - Interactive build-up of complex work. Set a variable, inspect it, reshape it, feed it back into the next cmdlet. Build a multi-step workflow one command at a time with every previous step's result still in scope.
# Command 1 — cold import, paid once for the whole session
Import-Module Az.Compute, Az.Storage, Az.Network
# Command 2 — instant; capture the result into a variable
$vms = Get-AzVM -Status
# Command 3 — instant; same session, $vms still in scope
$vms | Where-Object PowerState -eq "VM running" |
Group-Object Location | Sort-Object Count -Descending
# Command 4 — instant; reach into a new service, same session
Get-AzStorageAccount | Where-Object { -not $_.EnableHttpsTrafficOnly }PowerShell on SplashShell is the difference between "AI can answer one-off questions" and "AI can do real infrastructure work." bash and cmd are fully supported too, but pwsh is where SplashShell shines.
SplashShell opens a real, visible terminal window. You see every AI command as it runs — same characters, same output, same prompt — and you can type into the same window yourself at any time. When a command hangs on an interactive prompt, stalls in watch mode, or just needs a Ctrl+C, the AI can read what's currently on the screen and send keystrokes (Enter, y/n, arrow keys, Ctrl+C) back to the running command — diagnosing and responding without human intervention.
| Tool | Description |
|---|---|
start_console |
Open a visible terminal window. Pick a shell (bash, pwsh, powershell, cmd, or a full path). Optional cwd, banner, and reason parameters. Reuses an existing standby of the same shell unless reason is provided. |
execute_command |
Run a pipeline. Optionally specify shell to target a specific shell type — finds an existing console of that shell, or auto-starts one. Times out cleanly with output cached for wait_for_completion. On timeout, includes a partialOutput snapshot so the AI can diagnose stuck commands immediately. |
wait_for_completion |
Block until busy consoles finish and retrieve cached output (use after a command times out). |
peek_console |
Read-only snapshot of what a console is currently displaying. On Windows, reads the console screen buffer directly (exact match with the visible terminal). On Linux/macOS, uses a built-in VT terminal interpreter as fallback. Specify a console by display name or PID, or omit to peek at the active console. Reports busy/idle state, running command, and elapsed time. |
send_input |
Send raw keystrokes to a busy console's PTY input. Use \r for Enter, \x03 for Ctrl+C, \x1b[A for arrow up, etc. Rejected when the console is idle (use execute_command instead). Console must be specified explicitly — no implicit routing, for safety. Max 256 chars per call. |
Status lines include the console name, shell family, exit code, duration, and current directory:
✓ #12345 Sapphire (bash) | Status: Completed | Pipeline: ls /tmp | Duration: 0.6s | Location: /tmp
Claude Code–compatible file primitives (read_file, write_file, edit_file, search_files, find_files), useful when the MCP client doesn't already provide them.
Each console tracks its own cwd. When the active console is busy, the AI is auto-routed to a sibling console of the same shell family — started at the source console's cwd — and its next command runs immediately. Manual cd in the terminal is detected and the AI is warned before it runs the wrong command in the wrong place.
Other niceties: console re-claim — consoles outlive their parent MCP process, so AI client restarts don't kill your loaded modules and variables. Multi-line PowerShell — heredocs, foreach, try/catch, and nested scriptblocks all work. Sub-agent isolation — parallel AI agents each get their own consoles so they don't clobber each other's shells.
Full routing matrix
| Scenario | Behavior |
|---|---|
| First execute on a new shell | Auto-starts a console; warns so you can verify cwd before re-executing |
| Active console matches requested shell | Runs immediately |
| Active console busy, same shell requested | Auto-starts a sibling console at the source console's cwd and runs immediately |
| Switch to a same-shell standby | Prepends cd preamble so the command runs in the source cwd, then executes |
| Switch to a different shell | Warns to confirm cwd (cross-shell path translation is not implemented) |
User manually cd'd in the active console |
Warns so the AI can verify the new cwd before running its next command |
SplashShell runs as a stdio MCP server. When the AI calls start_console, SplashShell spawns itself in --console mode as a ConPTY worker, which hosts the actual shell (cmd.exe, pwsh.exe, bash.exe) inside a real Windows console window. The parent process streams stdin/stdout over a named pipe, injects OSC 633 shell integration scripts (the same protocol VS Code uses) to emit explicit command-lifecycle markers, and parses those markers to delimit command output, track cwd, and capture exit codes — no output-silence heuristics, no prompt-string detection.
Architecture diagram
graph TB
Client["MCP Client<br/>(Claude Code, etc.)"]
subgraph Proxy["SplashShell proxy (stdio MCP server)"]
CM["Console Manager<br/>(cwd tracking, re-claim,<br/>cache drain, switching)"]
Tools["start_console<br/>execute_command<br/>wait_for_completion<br/>peek_console / send_input<br/>read_file / write_file / edit_file<br/>search_files / find_files"]
end
subgraph Consoles["Visible Console Windows (each runs splash --console)"]
subgraph C1["#9876 Sapphire (bash)"]
PTY1["ConPTY + bash<br/>(+ OSC 633)"]
end
subgraph C2["#5432 Cobalt (pwsh)"]
PTY2["ConPTY + pwsh<br/>(+ OSC 633)"]
end
subgraph C3["#1234 Topaz (cmd)"]
PTY3["ConPTY + cmd.exe<br/>(+ OSC 633 via PROMPT)"]
end
end
User["User"] -- "keyboard" --> C1
User -- "keyboard" --> C2
User -- "keyboard" --> C3
Client -- "stdio" --> Proxy
CM -- "Named Pipe" --> C1
CM -- "Named Pipe" --> C2
CM -- "Named Pipe" --> C3
CM -. "auto-switch<br/>if busy" .-> C1
CM -. "shell routing" .-> C2
git clone https://github.com/yotsuda/SplashShell.git
cd SplashShell
dotnet publish -c Release -r win-x64 --no-self-contained -o ./distThe binary is ./dist/splash.exe. Use the absolute path instead of the npx command in your MCP config.
Windows is the primary target (ConPTY + Named Pipe, fully tested). Unix PTY fallback for Linux/macOS is experimental.
- cmd.exe exit codes always read as 0 — cmd's
PROMPTcan't expand%ERRORLEVEL%at display time, so AI commands show asFinished (exit code unavailable). Usepwshorbashfor exit-code-aware work. - Don't
Remove-Module PSReadLine -Forceinside a pwsh session — PSReadLine's background reader threads survive module unload and steal console input, hanging the next AI command. Not recoverable.
MIT. Release notes in CHANGELOG.md.
