io.github.mkpvishnu/terminal-mcp

Why This Exists

If you've hit any of these limitations with Claude Code, terminal-mcp solves them:

• "Claude Code can't handle interactive sessions" — The built-in Bash tool runs each command in a fresh subprocess. No persistence, no back-and-forth.
• "SSH not supported in Claude Code" — You can't SSH into a server and run multiple commands across an active connection.
• "Claude Code Bash tool doesn't support REPLs" — Python, Node, Ruby, and other interpreters need a persistent session for multi-line interaction.
• "How to use psql / mysql / redis-cli with Claude Code" — Database CLIs require a live connection that survives across tool calls.
• "Interactive terminal not working in Claude Code" — TUI apps (htop, vim, ncdu, fzf) need a real PTY with special key support.
• "Claude Code can't send arrow keys or Tab" — The Bash tool has no concept of terminal escape sequences.

terminal-mcp fills this gap by exposing MCP tools that create and manage real PTY sessions. Each session runs as a persistent child process; you send input, special keys, and control characters and read output across multiple tool calls for as long as the session lives.

Features

• Persistent PTY sessions — real terminal sessions that survive across tool calls
• Send + read in one call — session_interact combines send and read, halving LLM round trips
• Regex-triggered reads — session_wait_for blocks until a pattern matches in output — no more guessing timeouts
• Dangerous command gate — detects risky commands (rm -rf, DROP TABLE, curl|sh, etc.) and requires confirmation
• OSC 133 shell integration — auto-detects command boundaries and exit codes from modern shells
• Special key support — arrow keys, Tab, Escape, function keys (F1–F12), Home/End, Page Up/Down
• Control characters — Ctrl-C, Ctrl-D, Ctrl-Z, Ctrl-L, and telnet escape
• Four read modes — stream (waits for output to settle), snapshot (pyte screen buffer), auto (auto-detects TUI apps), and diff (returns only changed screen lines)
• Auto TUI detection — automatically detects alternate screen buffer (htop, vim, etc.) and switches to snapshot mode
• Output diff mode — returns only changed lines since last read, minimizing tokens for TUI monitoring
• Intelligent truncation — four truncation modes: tail (default), head_tail (preserves beginning and end), tail_only (for build logs), none
• ANSI stripping — optional removal of escape sequences for clean text output
• Idle cleanup — automatic session cleanup after configurable timeout
• Session management — list, label, and manage multiple concurrent sessions
• Dynamic resize — resize terminal dimensions on the fly with SIGWINCH support
• Secret input — send passwords without logging them
• Scrollback history — access terminal scrollback buffer beyond the visible screen
• One-shot execution — run a single command without manual session management
• Smart output truncation — four truncation strategies (tail, head_tail, tail_only, none) to prevent context overflow while preserving the most useful output
• Env var configuration — configure all settings via TERMINAL_MCP_* environment variables
• PyPI distribution — install directly with pip install terminal-mcp

Supported Clients

Quickstart

Install

Recommended — no install needed:

uvx terminal-mcp

Or install via pip:

pip install terminal-mcp

Or from source:

git clone https://github.com/mkpvishnu/terminal-mcp.git
cd terminal-mcp
pip install -e ".[dev]"

Register with Claude Code

Add to ~/.claude.json (or project .mcp.json):

{
  "mcpServers": {
    "terminal": {
      "command": "uvx",
      "args": ["terminal-mcp"]
    }
  }
}

Register with Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "terminal": {
      "command": "uvx",
      "args": ["terminal-mcp"]
    }
  }
}

Register with VS Code / Cursor

Click the one-click install badge above, or add to .vscode/mcp.json:

{
  "servers": {
    "terminal-mcp": {
      "command": "uvx",
      "args": ["terminal-mcp"]
    }
  }
}

Verify it works

session_exec  exec="echo hello from terminal-mcp"

Demo

session_create  command="ssh user@myserver.example.com"  label="prod-ssh"
session_read    session_id="a1b2c3d4"  timeout=5.0
session_send    session_id="a1b2c3d4"  password="mypassword"
session_send    session_id="a1b2c3d4"  input="df -h"
session_read    session_id="a1b2c3d4"
session_close   session_id="a1b2c3d4"

session_create  command="python3"  label="repl"
session_read    session_id="e5f6g7h8"
session_send    session_id="e5f6g7h8"  input="import math"
session_send    session_id="e5f6g7h8"  input="print(math.sqrt(144))"
session_read    session_id="e5f6g7h8"
session_close   session_id="e5f6g7h8"

session_create  command="python3 -m openclaw configure"  label="openclaw"
session_read    session_id="x1y2z3w4"  timeout=3.0
session_send    session_id="x1y2z3w4"  key="down"
session_send    session_id="x1y2z3w4"  key="down"
session_send    session_id="x1y2z3w4"  key="enter"
session_read    session_id="x1y2z3w4"
session_send    session_id="x1y2z3w4"  key="tab"
session_read    session_id="x1y2z3w4"
session_close   session_id="x1y2z3w4"

session_create   command="htop"  label="monitor"
session_read     session_id="a1b2c3d4"
→ auto-detects TUI, returns snapshot with mode_used="snapshot", tui_active=true

session_read     session_id="a1b2c3d4"  mode="diff"
→ returns only changed lines since last read

session_read     session_id="a1b2c3d4"  mode="diff"
→ returns only lines that changed, minimizing tokens

session_close    session_id="a1b2c3d4"

session_exec  exec="ls -la /tmp"
session_exec  exec="python3 -c 'print(42)'"  command="bash"  timeout=10.0

session_create   command="bash"  label="demo"
session_interact session_id="a1b2c3d4"  input="ls -la"  wait_for="\\$\\s*$"  timeout=5.0
session_interact session_id="a1b2c3d4"  input="whoami"  wait_for="\\$"
session_close    session_id="a1b2c3d4"

session_create   command="bash"  label="build"
session_send     session_id="a1b2c3d4"  input="npm run build"
session_wait_for session_id="a1b2c3d4"  pattern="Build complete|ERROR"  timeout=60.0
session_close    session_id="a1b2c3d4"

session_send    session_id="a1b2c3d4"  input="rm -rf /tmp/old"
→ returns: requires_confirmation=true, reason="Matched dangerous pattern: ..."

session_send    session_id="a1b2c3d4"  input="rm -rf /tmp/old"  confirmed=true
→ executes the command

session_send    session_id="a1b2c3d4"  control_char="c"
session_read    session_id="a1b2c3d4"

Tool Reference

session_create

Spawn a persistent PTY terminal session.

Returns: session_id, label, pid, created_at, snapshot_available

session_send

Send input text, a control character, or a special key to an active session. Only one of input, control_char, key, or password may be provided per call.

Returns: bytes_sent — or requires_confirmation, reason if the command matches a dangerous pattern

session_resize

Resize the terminal window of an active session.

Returns: rows, cols

session_read

Read output from a session.

Returns: output, bytes_read, prompt_detected, is_alive, truncated, tui_active, snapshot_available, mode_used, changed_lines (diff mode), is_first_read (diff mode), total_lines (scrollback), osc133, command_state, exit_code, command_complete (shell integration)

session_close

Terminate a session gracefully (EOF → SIGHUP → SIGKILL).

Returns: exit_status — or already_closed: true if the session was already terminated (idempotent)

session_exec

Execute a command in a temporary session and return output. The session is automatically cleaned up.

Returns: output, bytes_read, session_id, truncated

session_interact

Send input and read output in a single call. Combines session_send + session_read to halve round trips. Optionally waits for a regex pattern in the output.

Returns: output, bytes_read, bytes_sent, matched (when wait_for used), prompt_detected, is_alive, truncated, tui_active, mode_used, snapshot_available

session_wait_for

Read output from a session until a regex pattern matches or timeout expires. Use this instead of session_read when you know what output to expect.

Returns: output, bytes_read, matched, prompt_detected, is_alive, truncated

session_list

List all active sessions with their status and idle time.

Returns: sessions (array with tui_active, snapshot_available per session), count

Architecture

flowchart LR
    Client[AI Client] -->|MCP JSON-RPC| Server[terminal-mcp]
    Server --> SM[Session Manager]
    SM --> S1[PTY 1: bash]
    SM --> S2[PTY 2: python3]
    SM --> S3[PTY 3: ssh user@host]
    S1 & S2 & S3 -.->|PTY output| Reader[Reader Thread]
    Reader -.->|buffer| Server

stateDiagram-v2
    [*] --> Active : session_create
    state Active {
        Idle --> Sending : session_send
        Sending --> Idle
        Idle --> Reading : session_read
        Reading --> Idle
        Idle --> Resizing : session_resize
        Resizing --> Idle
    }
    Active --> [*] : session_close
    Active --> [*] : idle_timeout

Each session is backed by a real PTY allocated via pexpect.spawn. The design has four main parts:

Background reader thread. A daemon thread continuously reads from the PTY file descriptor in 4096-byte chunks and appends bytes to an in-memory buffer. The thread is lock-protected and dies automatically when the child process exits.

Output settling (stream mode). session_read in stream mode polls the buffer until no new bytes have arrived for timeout seconds (default 2s), then returns everything written since the last read call. A hard ceiling of timeout + 10s prevents infinite blocking.

Snapshot mode (always on). All PTY output is fed into a pyte virtual screen buffer. In auto mode (the default), session_read automatically detects TUI applications via alternate screen buffer sequences and returns a rendered snapshot. The diff mode returns only changed lines since the last read, minimizing tokens for TUI monitoring.

Idle cleanup. SessionManager runs a background cleanup loop (every 60s by default) that closes sessions idle longer than their idle_timeout. The default timeout is 30 minutes. Concurrent sessions are capped at 10 by default.

Configuration

All settings can be overridden via environment variables prefixed with TERMINAL_MCP_:

Per-session parameters: rows, cols, idle_timeout, and scrollback_lines can be set per session via session_create. There is no global env var for scrollback_lines — it defaults to 1000 lines per session. Set scrollback_lines=0 to disable scrollback history.

How to Set Environment Variables

Since terminal-mcp runs as a subprocess of your MCP client, environment variables must be configured in the client's MCP server configuration — not in your shell profile (.bashrc, .zshrc, etc.).

Claude Code (~/.claude.json or project .mcp.json):

{
  "mcpServers": {
    "terminal": {
      "command": "uvx",
      "args": ["terminal-mcp"],
      "env": {
        "TERMINAL_MCP_MAX_SESSIONS": "20",
        "TERMINAL_MCP_IDLE_TIMEOUT": "3600",
        "TERMINAL_MCP_SAFETY_GATE": "off",
        "TERMINAL_MCP_TRUNCATION_MODE": "head_tail"
      }
    }
  }
}

Claude Desktop (claude_desktop_config.json):

{
  "mcpServers": {
    "terminal": {
      "command": "uvx",
      "args": ["terminal-mcp"],
      "env": {
        "TERMINAL_MCP_MAX_SESSIONS": "20",
        "TERMINAL_MCP_IDLE_TIMEOUT": "3600"
      }
    }
  }
}

VS Code / Cursor (.vscode/mcp.json):

{
  "servers": {
    "terminal-mcp": {
      "command": "uvx",
      "args": ["terminal-mcp"],
      "env": {
        "TERMINAL_MCP_MAX_SESSIONS": "20",
        "TERMINAL_MCP_IDLE_TIMEOUT": "3600"
      }
    }
  }
}

Windsurf (~/.codeium/windsurf/mcp_config.json):

{
  "mcpServers": {
    "terminal": {
      "command": "uvx",
      "args": ["terminal-mcp"],
      "env": {
        "TERMINAL_MCP_MAX_SESSIONS": "20"
      }
    }
  }
}

You can also set env vars directly in your shell when running the server manually for testing:

TERMINAL_MCP_MAX_SESSIONS=20 TERMINAL_MCP_SAFETY_GATE=off uvx terminal-mcp

Changelog

v0.4.6

• Bump actions/github-script from 7 to 8 (#14)

v0.4.5

• Remove stale publish.yml workflow (#13)

v0.4.4

• Fix release push: use RELEASE_TOKEN for branch protection bypass (#12)

v0.4.3

• Fix wait_for matching command echo — session_interact with wait_for no longer matches the echoed input text. Pattern matching now starts from a buffer position anchored before the send, so only genuinely new output is matched (fixes #6)
• Fix wait_for matching stale buffer content — session_interact with wait_for no longer matches pre-existing unread buffer content (startup banners, previous command output). Uses a monotonic absolute byte counter that survives buffer trims (fixes #7)
• Idempotent session_close — closing an already-closed or naturally-exited session now returns success: true, already_closed: true instead of a not_found error. Safe for finally-style cleanup (fixes #8)
• Comprehensive ANSI stripping — strip_ansi now handles Kitty keyboard protocol sequences, application keypad mode (ESC=/ESC>), DCS sequences, secondary DA responses, and tilde-terminated CSI sequences that leaked through during TUI exit transitions (fixes #9)

v0.4.2

• Fix is_alive race on Linux — is_alive property now uses exception-safe _is_alive() internally, preventing PtyProcessError when child processes exit before waitpid can reap them. Fixes flaky CI failures on Linux runners

v0.4.1

• Auto TUI detection — automatically detects alternate screen buffer sequences (ESC[?1049h, ESC[?47h, ESC[?1047h) and switches session_read to snapshot mode. New mode="auto" is now the default
• Output diff mode — session_read with mode="diff" returns only changed screen lines since last read, with 1-indexed line numbers and is_first_read flag
• Intelligent truncation — new truncate_output_smart() with four strategies: tail (keep beginning), head_tail (keep first 30% + last 70% with line-count marker), tail_only (keep end, ideal for build logs), none (disable truncation). Configurable via truncation parameter on all read tools or TERMINAL_MCP_TRUNCATION_MODE env var
• Always-on pyte — snapshot mode is now always initialized (no need for enable_snapshot=true). enable_snapshot parameter deprecated
• read_mode on session_interact — choose how output is read back: auto, stream, snapshot, or diff
• Thread-safe screen reads — read_snapshot() and read_diff() now acquire buffer lock before reading pyte screen
• Response fields: tui_active, snapshot_available, mode_used added to read responses; snapshot_available added to create and list responses

v0.4.0

• session_interact tool — send input and read output in a single MCP call, halving round trips. Supports all input types (text, keys, control chars, passwords) with optional wait_for regex pattern
• session_wait_for tool — block until a regex pattern matches in session output or timeout expires. Replaces fragile timeout-based reads when you know what to expect
• Dangerous command gate — detects risky commands (rm -rf, DROP TABLE, curl|sh, chmod 777, etc.) and returns requires_confirmation instead of executing. Resend with confirmed=true to proceed. 17 built-in patterns, extensible via TERMINAL_MCP_DANGEROUS_PATTERNS env var, disable with TERMINAL_MCP_SAFETY_GATE=off
• OSC 133 shell integration — auto-detects command boundary markers emitted by modern shells (bash 5.2+, zsh, fish). When detected, read responses include command_state, exit_code, and command_complete fields for precise command completion tracking

v0.3.3

• Buffer memory cap — per-session PTY buffer capped at 1MB (configurable via TERMINAL_MCP_MAX_BUFFER_BYTES), prevents unbounded memory growth on long-running sessions
• Async event loop — all blocking PTY calls wrapped in asyncio.to_thread(), unblocking the event loop for concurrent MCP requests
• Snapshot ANSI stripping — strip_ansi parameter now correctly applied in snapshot and scrollback read modes
• Exec output truncation — session_exec now applies max_output_bytes truncation to prevent context overflow
• SIGTERM cleanup — added signal handler to close all PTY sessions on SIGTERM (Docker stop, kill, systemd)
• Close race condition — close() now handles pexpect exceptions when child process is already reaped
• Removed unused import atexit from server.py

v0.3.1

• MCP registry publication — added mcp-name marker and server.json for official MCP registry
• Version bump for registry metadata

v0.3.0

• Output truncation — large outputs are now automatically truncated to max_output_bytes (100KB default)
• Environment variable config — all settings configurable via TERMINAL_MCP_* env vars
• session_resize tool — dynamically resize terminal dimensions (sends SIGWINCH)
• Secret input — password parameter on session_send for credentials (redacted from logs)
• Scrollback buffer — pyte.HistoryScreen with configurable history depth; scrollback param on session_read
• session_exec tool — one-shot command execution with automatic session cleanup
• PyPI publishing — pip install terminal-mcp via trusted publishing workflow

v0.2.0

• Special key support — arrow keys, Tab, Escape, function keys (F1–F12), Home/End, Page Up/Down, and more via the key parameter on session_send
• Mutual exclusivity — input, control_char, and key are now validated as mutually exclusive
• Added GitHub Actions CI (Python 3.10–3.13) and CodeQL security scanning
• Added project metadata, classifiers, and MIT license

v0.1.0

• Initial release
• Persistent PTY sessions via pexpect
• Stream and snapshot read modes
• Control character support (Ctrl-C, Ctrl-D, Ctrl-Z, Ctrl-L)
• Session management with idle cleanup

Running Tests

pip install -e ".[dev]"
pytest tests/ -v

Contributing

Contributions are welcome! Please open an issue first to discuss what you'd like to change.

1. Fork the repository
2. Create a feature branch
3. Add tests for new functionality
4. Ensure all tests pass: pytest tests/ -v
5. Submit a pull request

License

MIT