terminal-mcp
A terminal emulator exposed via MCP for AI assistants
Documentation
Let AI see and interact with your terminal.
Terminal MCP gives LLMs a shared view of your terminal session. Perfect for debugging CLIs and TUI applications in real-time, or letting AI drive terminal-based tools autonomously.
Install
npm install -g @ellery/terminal-mcpOr via install script:
curl -fsSL https://raw.githubusercontent.com/elleryfamilia/terminal-mcp/main/install.sh | bashConfigure your AI tools
Wire `terminal-mcp` into the MCP config of every AI tool installed on your machine in one shot:
terminal-mcp setup # detect & install for all detected tools
terminal-mcp setup --dry-run # preview without writing
terminal-mcp setup --client claude-code,gemini # specific tools only
terminal-mcp setup --uninstall # remove the entry from each toolSupported clients (each gets the right schema for its config format):
| Client | Config file | Format |
|---|---|---|
| OpenAI Codex CLI | `~/.codex/config.toml` | TOML |
| GitHub Copilot CLI | `~/.copilot/mcp-config.json` | JSON |
| Gemini CLI | `~/.gemini/settings.json` | JSON |
| OpenCode | `~/.config/opencode/opencode.json` | JSON |
| Claude Code | `~/.claude.json` | JSON |
| Claude Desktop | `~/Library/Application Support/Claude/claude_desktop_config.json` (macOS) / `%APPDATA%\Claude\claude_desktop_config.json` (Windows) | JSON |
A `.bak` of any pre-existing config is written next to the original on first install. The `terminal-mcp` entry is added without disturbing other servers or unrelated keys; running `setup` again is a no-op.
Upgrading
npm install -g @ellery/terminal-mcp@latestInteractive mode will print a banner on next launch when a newer release is available — `terminal-mcp` checks the npm registry once per day and caches the result. Headless and MCP-client modes never check or print anything (so MCP stdio stays clean). To opt out entirely, set `NO_UPDATE_NOTIFIER=1` or pass `--no-update-notifier`.
Features
- Full Terminal Emulation: Uses xterm.js headless for accurate VT100/ANSI emulation
- Cross-Platform PTY: Native pseudo-terminal support via node-pty (macOS, Linux, Windows)
- MCP Protocol: Implements Model Context Protocol for AI assistant integration
- Session Recording: Record terminal sessions to asciicast format for playback with asciinema
- Simple API: Nine tools covering input, observation, recording, and session lifecycle
- Headless Mode: Run as a standalone MCP server without a TTY — ideal for CI, containers, and non-interactive environments
- Multi-Session: Run multiple isolated terminal sessions in one process, addressed by `sessionId`
- Sandbox Mode: Optional security restrictions for filesystem and network access
Building from Source
npm install
npm run buildUsage
MCP Configuration
Add to your MCP client settings:
{
"mcpServers": {
"terminal": {
"command": "terminal-mcp"
}
}
}With custom options:
{
"mcpServers": {
"terminal": {
"command": "terminal-mcp",
"args": ["--cols", "100", "--rows", "30", "--shell", "/bin/zsh"]
}
}
}Command-Line Options
terminal-mcp [OPTIONS]
Options:
--cols Terminal width in columns (default: 120)
--rows Terminal height in rows (default: 40)
--shell Shell to use (default: $SHELL or bash)
--headless Run in headless mode (embedded PTY + MCP over stdio, no TTY needed)
--sandbox Enable sandbox mode (restricts filesystem/network)
--sandbox-config Load sandbox config from JSON file
--version, -v Show version number
--help, -h Show help message
Recording Options:
--record [mode] Enable recording (default mode: always)
Modes: always, on-failure, off
--record-dir Recording output directory
(default: ~/.local/state/terminal-mcp/recordings)
--idle-time-limit Max idle time between events (default: 2s)
--max-duration Max recording duration (default: 3600s)
--inactivity-timeout Stop after no output (default: 600s)
Multi-Session Options:
--max-sessions Max concurrent sessions (default: 5)
--session-idle-timeout Idle non-default sessions are auto-destroyed
after this period (default: 600s)Headless Mode
By default, Terminal MCP uses a dual-process architecture: you run `terminal-mcp` in an interactive terminal (which creates a Unix socket), then your MCP client spawns a second instance that connects to that socket. This requires a TTY.
Headless mode (`--headless`) eliminates this requirement by spawning an embedded PTY internally and serving MCP directly over stdio in a single process. No interactive terminal session, no socket — just a self-contained MCP server with a built-in terminal.
When to use headless mode
- CI/CD pipelines — no TTY available
- Docker containers — no interactive shell to run alongside
- Remote/cloud environments — MCP servers spawned by automation
- Simplified setup — single process, no socket coordination needed
Configuration
{
"mcpServers": {
"terminal": {
"command": "terminal-mcp",
"args": ["--headless", "--cols", "120", "--rows", "40"]
}
}
}How it works
MCP Client (Claude Code, etc.)
│ STDIO (JSON-RPC)
▼
terminal-mcp --headless
├── MCP Server (stdio transport)
├── Terminal Emulator (@xterm/headless)
└── Embedded PTY (node-pty)
│
▼
Shell Process (bash, zsh, etc.)In headless mode, the terminal session is initialized eagerly at startup, so all tools (`type`, `sendKey`, `getContent`, `takeScreenshot`, `startRecording`, `stopRecording`, `createSession`, `listSessions`, `destroySession`) are available immediately.
MCP Tools
All input/output tools (`type`, `sendKey`, `getContent`, `takeScreenshot`) accept an optional `sessionId` argument. Omit it to target the default session; pass the ID returned by `createSession` to drive a specific session.
`type`
Send text input to the terminal.
{
"name": "type",
"arguments": {
"text": "echo hello"
}
}`sendKey`
Send special keys or key combinations.
{
"name": "sendKey",
"arguments": {
"key": "Enter"
}
}Supported keys:
- Basic: `Enter`, `Tab`, `Escape`, `Backspace`, `Delete`
- Arrow: `ArrowUp`, `ArrowDown`, `ArrowLeft`, `ArrowRight`
- Navigation: `Home`, `End`, `PageUp`, `PageDown`, `Insert`
- Function: `F1` through `F12`
- Control: `Ctrl+A` through `Ctrl+Z`, `Ctrl+C`, `Ctrl+D`, etc.
`getContent`
Get the terminal buffer as plain text.
{
"name": "getContent",
"arguments": {
"visibleOnly": false
}
}`takeScreenshot`
Capture the terminal state. Supports three output formats:
| Format | Description |
|---|---|
| `text` (default) | JSON with plain text content, cursor position, and dimensions |
| `ansi` | JSON with ANSI color escape codes preserved in the content field |
| `png` | Color screenshot as a PNG image (requires `@resvg/resvg-js`) |
{
"name": "takeScreenshot",
"arguments": { "format": "text" }
}The `ansi` format reconstructs SGR escape sequences from the terminal's cell buffer, preserving 16-color, 256-color, and 24-bit truecolor attributes along with bold, dim, italic, and underline styles.
The `png` format returns an MCP `image` content block with base64-encoded PNG data, rendered with the One Dark color theme and macOS-style window chrome.
`startRecording`
Start recording terminal output to an asciicast v2 file.
{
"name": "startRecording",
"arguments": {
"mode": "always",
"idleTimeLimit": 2,
"maxDuration": 3600
}
}Options:
- `mode`: `always` (save all) or `on-failure` (save only on non-zero exit)
- `outputDir`: Custom output directory
- `idleTimeLimit`: Max seconds between events (caps pauses in playback)
- `maxDuration`: Auto-stop after N seconds
- `inactivityTimeout`: Auto-stop after N seconds of no output
`stopRecording`
Stop a recording and finalize the asciicast file.
{
"name": "stopRecording",
"arguments": {
"recordingId": "abc123"
}
}`createSession`
Create a new terminal session and return its metadata. Use the returned `sessionId` to target this session in subsequent tool calls.
{
"name": "createSession",
"arguments": {
"shell": "/bin/zsh",
"cols": 100,
"rows": 30
}
}All arguments are optional. Returns:
{
"sessionId": "3029d",
"shell": "/bin/zsh",
"cols": 100,
"rows": 30,
"createdAt": "2026-04-25T12:58:01.072Z",
"lastActivityAt": "2026-04-25T12:58:01.072Z",
"isDefault": false
}`listSessions`
List all active sessions including the default. Reports configured limits.
{ "name": "listSessions", "arguments": {} }`destroySession`
Destroy a session by ID. The default session cannot be destroyed.
{
"name": "destroySession",
"arguments": { "sessionId": "3029d" }
}Multi-Session
By default, every tool call without a `sessionId` targets a single auto-created default session — the same behavior the project has always had. Pass `sessionId` to drive multiple isolated PTYs from one process.
- The default session is created on first use and cannot be destroyed.
- Additional sessions are created by `createSession` and tracked until they're destroyed or idle-evicted (`--session-idle-timeout`, default 600s).
- Concurrent sessions are capped at `--max-sessions` (default 5).
- An active recording captures output from all sessions in the process.
Typical use case: an AI agent driving a long-running build in one session while running diagnostics in another, without command interleaving.
Sandbox Mode
Run the terminal with restricted filesystem and network access:
# Interactive permission configuration
terminal-mcp --sandbox
# With a config file
terminal-mcp --sandbox --sandbox-config ~/.terminal-mcp-sandbox.jsonThe interactive mode shows a TUI dialog to configure permissions:
- Read/Write: Full access (current directory, /tmp, caches)
- Read-Only: Can read but not modify (home directory)
- Blocked: No access (SSH keys, cloud credentials, auth tokens)
Example config file:
{
"filesystem": {
"readWrite": [".", "/tmp", "~/.cache"],
"readOnly": ["~"],
"blocked": ["~/.ssh", "~/.aws", "~/.gnupg"]
},
"network": {
"mode": "all"
}
}Platform support:
- macOS: Full support via sandbox-exec (Seatbelt)
- Linux: Full support via bubblewrap (requires `bwrap` installed)
- Windows: Graceful fallback (runs without sandbox)
See Sandbox Documentation for detailed configuration options.
Recording
Terminal MCP can record sessions to asciicast v2 format, compatible with asciinema for playback.
Quick Start
# Start with recording enabled
terminal-mcp --record
# Run your commands, then exit
exit
# Output shows the saved file path:
# Recordings saved:
# ~/.local/state/terminal-mcp/recordings/20240115_143022.cast
#
# Play with: asciinema playPlayback
Install asciinema to play back recordings:
# macOS
brew install asciinema
# Linux/pip
pip install asciinema
# Play a recording
asciinema play ~/.local/state/terminal-mcp/recordings/20240115_143022.cast
# Play at 2x speed
asciinema play -s 2 recording.castRecording Modes
- `always` (default): Save every recording
- `on-failure`: Only save if the session exits with a non-zero code (useful for debugging failed CI runs)
# Only save recordings when something fails
terminal-mcp --record=on-failureMCP Tool Recording
AI assistants can also control recording programmatically via MCP tools:
1. Call `startRecording` to begin capturing
2. Perform terminal operations
3. Call `stopRecording` to finalize and save
This enables AI-driven workflows like "record this debugging session" or "capture this demo".
Architecture
Terminal MCP has three operating modes:
| Mode | Flag | Stdin | Description |
|---|---|---|---|
| Interactive | *(default)* | TTY | User gets a shell; AI connects via Unix socket |
| Client | *(default)* | non-TTY | Connects to an interactive session's socket, serves MCP over stdio |
| Headless | `--headless` | any | Self-contained: embedded PTY + MCP server over stdio |
Headless mode (recommended for MCP configs)
MCP Client (Claude Code, etc.)
│ STDIO (JSON-RPC)
▼
terminal-mcp --headless
├── MCP SDK (@modelcontextprotocol/sdk)
├── Terminal Emulator (@xterm/headless)
└── Embedded PTY (node-pty)
│
▼
Shell Process (bash, zsh, etc.)Interactive + Client mode (two-process)
terminal-mcp (interactive, in your terminal)
├── User shell (stdin/stdout)
└── Unix socket server (/tmp/terminal-mcp.sock)
▲
│ JSON-RPC over socket
▼
terminal-mcp (client, spawned by MCP client)
└── MCP server (stdio transport)Example Session
# Type a command
{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"type","arguments":{"text":"ls -la"}}}
# Send Enter key
{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"sendKey","arguments":{"key":"Enter"}}}
# Get the output
{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"getContent","arguments":{}}}Development
npm run build # Compile TypeScript
npm run dev # Run with tsx (development)Documentation
See the docs folder for detailed documentation:
Requirements
- Node.js 18.0.0 or later
- Windows 10 version 1809 or later (for ConPTY support)
License
MIT
Frequently asked questions
What is terminal-mcp?
terminal-mcp is A terminal emulator exposed via MCP for AI assistants
How do I install terminal-mcp?
Open the GitHub repository and follow its README. Most MCP servers are added to your client's MCP config, then called by your agent.
Is terminal-mcp open source?
Yes — it is hosted on GitHub at https://github.com/elleryfamilia/terminal-mcp and has 134 stars.
Related MCP tools
Model Context Protocol Servers
The Open-Source Multimodal AI Agent Stack: Connecting Cutting-Edge AI Models and Agent Infra
A MCP for Claude Desktop / Claude Code / Windsurf / Cursor to build n8n workflows for you
MCP server to provide Figma layout information to AI coding agents like Cursor
The world's best AI personal assistant for email. Open source app to help you reach inbox zero fast.
Instant is the best backend for AI-coded apps. You get auth, permissions, storage, presence, and streams — everything you need to ship apps your users will love.
Run your own MCP server? See who uses it and what to fix.
Measure it with TrackMCP