iterm2-agent
MCP server for controlling iTerm2 terminal sessions, enabling screen reading, command execution, keystroke sending, and session management through natural language.
README
iterm2-agent
MCP server for controlling iTerm2 terminal sessions via the iTerm2 Python API.
Read the screen, run commands, send keystrokes, monitor output, and manage panes — all through Model Context Protocol.
Prerequisites
- macOS with iTerm2 installed
- iTerm2 Python API enabled: Preferences → General → Magic → Enable Python API
- Python 3.11+
- uv (
brew install uv)
Install
MCP Server
git clone git@github.com:xjthy001/iterm2-agent.git
cd iterm2-agent
uv venv
uv pip install -e .
Then register it in ~/.claude.json (see Usage with Claude Code below).
Skill (optional)
Install the agent skill via skills.sh:
npx skills add xjthy001/iterm2-agent
Or manually:
mkdir -p ~/.claude/skills/iterm2-agent
cp skills/iterm2-agent/SKILL.md ~/.claude/skills/iterm2-agent/SKILL.md
The skill gives your AI agent a reference guide for using the 6 tools effectively. It auto-activates when you mention iTerm2 or terminal control.
Tools
| Tool | Purpose |
|---|---|
read_screen |
Read visible terminal content and cursor position |
run_command |
Execute a shell command, wait for output to stabilize, return result |
send_text |
Send raw text to a session (with optional Enter) |
send_control |
Send control characters (Ctrl+C, Ctrl+Z, Ctrl+D, etc.) |
watch_output |
Monitor output until a regex pattern matches |
manage_session |
List, create, split, close, or focus sessions |
read_screen
Read the visible screen content of a session.
lines: int = -1 # Number of lines to read (-1 = all visible)
session_id: str = "" # Target session (empty = active session)
run_command
Execute a command and capture output. Waits for output to stabilize (2s idle) before returning.
command: str # Shell command to execute
timeout: int = 30 # Max seconds to wait
session_id: str = ""
Commands are classified by a security guard:
- SAFE — read-only commands (
ls,pwd,git status, ...) - CAUTION — modifying commands (
mkdir,npm install,git push, ...) - DANGEROUS — destructive commands (
rm,sudo,kill, ...) — produces a warning
send_text
Send text to a session without automatically pressing Enter. Set press_enter=true to submit.
text: str # Text to send
press_enter: bool = false
session_id: str = ""
send_control
Send a control character.
character: str # One of: C, Z, D, L, ESCAPE, A, E, U, K, W, R
session_id: str = ""
| Character | Key | Action |
|---|---|---|
C |
Ctrl+C | Interrupt |
Z |
Ctrl+Z | Suspend |
D |
Ctrl+D | EOF |
L |
Ctrl+L | Clear screen |
ESCAPE |
Esc | Escape |
A |
Ctrl+A | Beginning of line |
E |
Ctrl+E | End of line |
U |
Ctrl+U | Kill line |
K |
Ctrl+K | Kill to end of line |
W |
Ctrl+W | Kill word |
R |
Ctrl+R | Reverse search |
watch_output
Block until a regex pattern appears on screen, or timeout.
pattern: str # Regex pattern to match
timeout: int = 60 # Max seconds to wait
session_id: str = ""
manage_session
Manage iTerm2 sessions.
action: str # list | create | split | close | focus
session_id: str = "" # Required for close/focus, optional for split
direction: str = "horizontal" # horizontal | vertical (split only)
Usage with Claude Code
1. Register the MCP server
Add the iterm2-agent entry to the mcpServers object in ~/.claude.json:
{
"mcpServers": {
"iterm2-agent": {
"type": "stdio",
"command": "/path/to/iterm2-agent/.venv/bin/python",
"args": ["-m", "iterm2_agent"]
}
}
}
Replace /path/to/iterm2-agent with your actual clone path. The command must point to the Python binary inside the project's virtualenv so that dependencies are available.
If
~/.claude.jsonalready has amcpServersobject, just add the"iterm2-agent": { ... }entry alongside existing servers. Don't overwrite the whole file.
Restart Claude Code after editing. The server connects to iTerm2 on startup — make sure iTerm2 is running and the Python API is enabled.
2. Install the skill (optional)
npx skills add xjthy001/iterm2-agent
Or manually copy from the repo:
mkdir -p ~/.claude/skills/iterm2-agent
cp skills/iterm2-agent/SKILL.md ~/.claude/skills/iterm2-agent/SKILL.md
After installing, invoke it with /iterm2-agent or it auto-activates when you mention terminal/iTerm2.
3. Verify
Start a new Claude Code session and try:
- "What's on my terminal right now?"
- "Run
git statusin iTerm2" - "Split the terminal and run tests in the new pane"
- "Start the dev server and tell me when it's ready"
- "Send Ctrl+C to stop the running process"
Architecture
┌─────────────────────────────────────┐
│ MCP Client │
│ (Claude Code / Claude Desktop) │
└──────────────┬──────────────────────┘
│ stdio (JSON-RPC)
┌──────────────▼──────────────────────┐
│ iterm2-agent (MCP Server) │
│ │
│ server.py FastMCP + lifespan│
│ connection.py Session resolver │
│ security.py Command classifier│
│ tools/ │
│ read_screen.py │
│ run_command.py │
│ send_text.py │
│ send_control.py │
│ watch_output.py │
│ manage_session.py │
└──────────────┬──────────────────────┘
│ WebSocket
┌──────────────▼──────────────────────┐
│ iTerm2 Python API │
│ (iterm2.Connection) │
└──────────────┬──────────────────────┘
│
┌──────────────▼──────────────────────┐
│ iTerm2.app │
└─────────────────────────────────────┘
Project Structure
iterm2-agent/
├── pyproject.toml
├── config/
│ └── default.toml
├── src/iterm2_agent/
│ ├── __init__.py
│ ├── __main__.py # Entry point: python -m iterm2_agent
│ ├── server.py # FastMCP server with iTerm2 lifespan
│ ├── connection.py # iTerm2Context, session resolution, screen reading
│ ├── security.py # Command classification (SAFE/CAUTION/DANGEROUS)
│ └── tools/
│ ├── __init__.py # Tool registration
│ ├── read_screen.py
│ ├── run_command.py
│ ├── send_text.py
│ ├── send_control.py
│ ├── watch_output.py
│ └── manage_session.py
├── tests/
│ ├── test_security.py # Security guard unit tests
│ ├── test_read_screen.py # Session resolution unit tests
│ ├── test_run_command.py # Security integration tests
│ └── test_send_control.py # Control character mapping tests
├── test_integration.py # Live integration tests (requires iTerm2)
└── skills/iterm2-agent/
└── SKILL.md # Agent skill definition (skills.sh compatible)
Testing
Unit tests (no iTerm2 required):
uv pip install pytest pytest-asyncio
uv run pytest
Live integration tests (requires iTerm2 running):
uv run python test_integration.py
License
MIT
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。