powershell-terminal
Enables Claude to execute PowerShell commands and scripts in a persistent interactive terminal on Windows, with real-time output streaming to a web terminal.
README
<!-- mcp-name: io.github.TiM00R/powershell-terminal-mcp -->
PowerShell Terminal
Shared AI + user PowerShell session on Windows ā execute commands, run scripts, automate your PC
PowerShell Terminal lets Claude (the AI assistant) run commands in a persistent, interactive PowerShell 7 session on your Windows machine through a real pseudo-terminal (ConPTY). Watch every command stream into your browser in real time while Claude receives smart-filtered output optimized for token efficiency.
šÆ What Is This?
Imagine telling Claude:
"Check what Python version I have and install requests if it's missing"
"Run my build script and tell me if anything failed"
"Find all .log files modified today and show me any errors"
"Save this cleanup script and run it every time I ask"
And Claude does it ā executing commands in your real PowerShell session, analyzing output, saving reusable scripts, and taking action on your behalf.
That's PowerShell Terminal.
⨠Key Features
Core Capabilities
- š„ļø Real PowerShell Session ā Persistent
pwsh(PS7) or PS5.1 session via ConPTY; state carries across commands (working directory, variables, activated venv) - š Shared Human + AI Terminal ā NiceGUI + xterm.js web terminal at
http://localhost:8090; type your own commands alongside Claude's - š Multi-Terminal Sync ā Open multiple browser tabs, all perfectly synchronized
- šŖ No Popup Windows ā Native console executables (
git,python,ipconfig, full paths) run inside ConPTY without spawning new windows - āļø Dual-Stream Output ā You see full output in the browser; Claude receives a token-reduced summary
- ā Reliable Completion Detection ā Exit codes and command completion detected via invisible prompt token in OSC escape sequences ā no fragile regex matching
- āØļø Interactive Commands ā Commands that prompt for input (
Read-Host) work; Claude can send input and interrupt with Ctrl+C - š Script Library ā Save and reuse named
.ps1scripts; full output persisted on script runs - šļø Command History ā Commands grouped into conversations and logged to SQLite with selective output persistence
The Interactive Web Terminal
PowerShell Terminal provides a fully interactive terminal window in your browser at http://localhost:8090 ā it looks and feels just like a native PowerShell window:
You can:
- Type commands directly (just like any terminal)
- Right-click to Copy/Paste, or use Ctrl+Shift+C / Ctrl+Shift+V
- Scroll through the full session scrollback
- Watch every command Claude runs appear in real time
Claude can:
- Execute commands that stream into your terminal
- See results instantly
- Continue working while you watch
The key advantage: Complete visibility and control. Every command Claude runs appears in your terminal in real time. You're never in the dark ā it's like sitting side-by-side with an assistant who types commands while you watch the screen.
Multi-Terminal Support: Open multiple browser windows at http://localhost:8090 ā they all stay perfectly synchronized via WebSocket broadcast. Type in one terminal, see it in all terminals instantly.
The Dual-Stream Architecture
PowerShell Session Output (ConPTY)
|
[Raw Output]
|
---------+---------
| |
[FULL] [FILTERED]
| |
v v
Web Terminal Claude
(You see all) (Smart summary)
- You: Full output, colors, and scrollback in the browser terminal
- Claude: Token-efficient filtered summary
- Both: Same live PowerShell session, synchronized state
Native Exe ā No Popup Windows
A key problem with running an AI-controlled terminal on Windows: native console executables like python, git, ipconfig, or any .exe would spawn a separate conhost.exe popup window, breaking the in-terminal experience.
PowerShell Terminal solves this completely:
- A PostCommandLookupAction hook intercepts every native CUI executable before it runs
- A PE header check distinguishes console apps (CUI) from GUI apps ā GUI apps like
notepadandcodeopen normally without blocking - Execution is handled via
System.Diagnostics.ProcesswithCreateNoWindow=trueand redirected I/O, so output flows through ConPTY instead of a new window - Works for short names (
git,python), full paths (D:\tools\ffmpeg.exe), and any exe not known in advance
š Quick Start
Requirements
- Windows 10 1809+ or Windows 11 (ConPTY required)
- PowerShell 7 (
pwsh) recommended; falls back to Windows PowerShell 5.1 - Python 3.10+
Option A ā Install from PyPI
Step 1: Create a virtual environment
mkdir D:\powershell_terminal
cd D:\powershell_terminal
py -m venv .venv
.\.venv\Scripts\Activate.ps1
Step 2: Install the package
pip install powershell-terminal-mcp
Step 3: Register with Claude Desktop
notepad $env:APPDATA\Claude\claude_desktop_config.json
Add:
{
"mcpServers": {
"powershell-terminal": {
"command": "D:\\powershell_terminal\\.venv\\Scripts\\powershell-terminal-mcp.exe"
}
}
}
Option B ā Install from Source (dev)
Step 1: Clone the repo
git clone https://github.com/TiM00R/powershell-terminal-mcp D:\powershell_terminal
cd D:\powershell_terminal
Step 2: Create the virtual environment and install dependencies
.\setup_venv.ps1
This creates .venv, installs all dependencies (including pywinpty), and installs the project in editable mode.
Step 3: Register with Claude Desktop
notepad $env:APPDATA\Claude\claude_desktop_config.json
Add:
{
"mcpServers": {
"powershell-terminal": {
"command": "D:\\powershell_terminal\\.venv\\Scripts\\python.exe",
"args": ["D:\\powershell_terminal\\src\\mcp_server.py"]
}
}
}
Fully quit and relaunch Claude Desktop (system tray ā Exit), then open a new conversation. On the first tool call the server starts the PowerShell session and opens the web terminal at http://localhost:8090.
š” Usage Examples
Running Commands
"What Python version do I have?"
"Run git status in D:\myproject"
"List all files modified today in my Downloads folder"
"Check if port 8080 is in use"
Scripts & Automation
"Save this build script and run it"
"Run my cleanup script"
"List my saved scripts"
Diagnostics & System Info
"Check disk space on all drives"
"What processes are using the most memory?"
"Show me the last 20 lines of the application event log"
"Is the Docker daemon running?"
Development Workflows
"Activate the venv and run the tests"
"Install requests and check if it imports correctly"
"Run git log --oneline -10 and summarize what changed"
šØ Example Session
You:
Run my test suite and tell me what failed
Claude:
I'll run the tests now...
Web terminal opens automatically at http://localhost:8090
Web Terminal shows (real-time):
PS D:\myproject> python -m pytest tests/ -v
============================= test session starts ==============================
collected 42 items
tests/test_api.py::test_login PASSED [ 2%]
tests/test_api.py::test_auth_required PASSED [ 4%]
tests/test_db.py::test_connection FAILED [ 26%]
...
========================= 3 failed, 39 passed in 8.42s =========================
Claude receives (filtered summary):
pytest: 42 tests, 3 failed, 39 passed (8.4s)
Failed: test_connection, test_insert_duplicate, test_rollback
Claude responds:
3 tests failed, all in
test_db.pyā looks like a database connection issue. Want me to look at the error details and fix them?
šļø Architecture
How It Works
Claude <--stdio JSON-RPC--> mcp_server.py
|
SharedTerminalState
/ | \
PwshSession OutputBuffer SmartOutputFilter SQLite (db.py)
(ConPTY via (full human (AI token- (conversations,
pywinpty) scrollback) reduced view) commands, scripts)
|
NiceGUI + xterm.js web terminal (shared, multi-tab)
- A single
pwshprocess runs inside a ConPTY (pywinpty). A background thread reads its output into a circular buffer. - The session's prompt is overridden (in the MCP session only) to emit a unique completion token + exit status inside invisible OSC escape sequences. The browser swallows them; the server detects them in the raw stream.
- An output start-marker (emitted via a PSReadLine Enter handler, also invisible) lets the server separate a command's real output from the terminal's echo of the typed command.
- Human keystrokes from the browser are passed through raw and are not tracked by the AI's completion detection, so the two streams never collide.
Project Structure
powershell_terminal/
āāā config/
ā āāā config.yaml # Web port, filter thresholds, error patterns
āāā data/
ā āāā commands.db # SQLite: conversations, commands, scripts
āāā scripts/ # Headless test harnesses
ā āāā test_pwsh_session.py # Session: completion, exit codes, interactive, Ctrl+C
ā āāā test_session_output.py # Buffer + dual-stream filter
ā āāā test_mcp_dispatch.py # All MCP tools via direct dispatch
āāā src/
ā āāā config/ # Configuration loading
ā āāā output/ # Output filtering and buffering
ā āāā pwsh/ # PowerShell session (ConPTY)
ā ā āāā pwsh_launch.py # Shell spawn, init script, native exe hook
ā ā āāā pwsh_session.py # Session lifecycle, run_command, send_input
ā ā āāā session_output.py # Dual-stream wrapper (raw + filtered)
ā ā āāā completion_token.py # Prompt token and OSC escape injection
ā āāā web/
ā ā āāā web_terminal.py # NiceGUI + xterm.js web terminal
ā āāā db.py # SQLite database layer
ā āāā mcp_server.py # MCP server entry point (all tools)
ā āāā shared_state.py # Global session hub
āāā setup_venv.ps1 # One-command environment setup
āāā run_web.py # Launch web terminal standalone (no Claude)
Technology Stack
- Python 3.10+ ā Core language
- MCP Protocol ā Claude integration (stdio JSON-RPC)
- pywinpty ā ConPTY pseudo-terminal on Windows
- NiceGUI + WebSockets ā Web terminal with multi-tab sync
- SQLite ā Command history and script storage
- xterm.js ā Browser terminal renderer
š§ MCP Tools Reference
Terminal / Execution
| Tool | Description |
|---|---|
execute_command(command, timeout?) |
Run a command; returns filtered output + exit code. Returns status: "running" on timeout. |
get_command_output(command_id, raw?) |
Fetch a prior command's output by id. |
send_input(text) |
Answer a running interactive command, then wait for completion. |
send_interrupt() |
Send Ctrl+C to the running command. |
get_terminal_status() |
Session alive? Web terminal URL. |
restart_session() |
Kill and respawn the PowerShell session (clears all state). |
open_terminal() |
Open (or re-open) the web terminal in the browser. |
Scripts
| Tool | Description |
|---|---|
save_script(name, content) |
Save (or overwrite) a named .ps1 script. |
list_scripts() |
List all saved scripts. |
run_script(name, timeout?) |
Run a saved script; full output is always persisted. |
Conversations (History)
| Tool | Description |
|---|---|
start_conversation(label?) |
Group subsequent commands; returns conversation_id. |
end_conversation(conversation_id?, status?) |
End the active (or specified) conversation. |
list_conversations(limit?) |
List recent conversations. |
get_conversation_commands(conversation_id) |
Commands logged under a conversation. |
š§ Configuration
config.yaml (project root) controls:
server.host/server.portā Web terminal address (defaultlocalhost:8090)- Output filter thresholds and error patterns
- SQLite database lives at
data\commands.dbin the project root
š”ļø Security Considerations
- Web terminal bound to
localhostonly ā not exposed to the network - Full command audit trail in SQLite
- The init script runs in the MCP's own session only ā your
$PROFILE, normal PowerShell, and prompt are never modified - Claude runs commands in your local user context with your normal permissions
š Known Issues & Limitations
- Windows only ā ConPTY is a Windows API; Linux/Mac not supported
- Interactive TUI apps not supported ā Commands that take over the terminal (e.g.
vim,htop) will hang; use-NonInteractivealternatives - Single session ā One shared PowerShell session; no per-command isolation
š Development
Run the headless test harnesses without Claude Desktop:
.\.venv\Scripts\python.exe scripts\test_pwsh_session.py # session: completion, exit codes, interactive, Ctrl+C, restart
.\.venv\Scripts\python.exe scripts\test_session_output.py # buffer + dual-stream filter
.\.venv\Scripts\python.exe scripts\test_mcp_dispatch.py # all MCP tools via direct dispatch
python run_web.py # launch web terminal standalone
š Version History
v0.1.0 (July 2026) ā Initial public release
- ā
ConPTY-based persistent PowerShell 7 session via
pywinpty - ā Shared human + AI terminal: NiceGUI + xterm.js web terminal, multi-tab WebSocket sync
- ā
Native exe fix: all CUI executables run inside ConPTY without popup windows (
System.Diagnostics.Process+CreateNoWindow=true+ redirected I/O) - ā PostCommandLookupAction hook covers short names, full paths, and any unknown exe
- ā PE header subsystem check: GUI apps (notepad, code) bypass the hook and open normally
- ā PATHEXT fix: session always gets a correct PATHEXT so bare command names resolve reliably
- ā Dual-stream output: full output in browser, token-reduced view for Claude
- ā Reliable command completion via prompt token in invisible OSC escape sequences
- ā
Interactive commands:
Read-Host,Ctrl+C,send_input - ā SQLite history: conversations, commands (selective persistence), saved scripts
- ā Full MCP tool set: execute, send_input, interrupt, restart, scripts, conversations
š¤ Contributing
This is Tim's personal project. If you'd like to contribute:
- Test on your setup and document any issues found
- Suggest improvements or missing features
- Share useful scripts you create
š License
MIT
Ready to let Claude run PowerShell for you? Register the MCP server in Claude Desktop and open a new conversation to get started.
Version: 0.1.0 Last Updated: July 2026 Maintainer: Tim
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.