pty-mcp
Provides a pseudo-terminal (PTY) interface that allows AI agents to interact with command-line tools requiring interactive prompts. It enables agents to autonomously spawn processes, read output, and send inputs for workflows like database migrations and project scaffolding.
README
pty-mcp
MCP server that gives AI coding agents a real pseudo-terminal (PTY) for handling interactive CLI prompts. Spawn processes, read their output, and send intelligent responses — no human in the loop.
Why I Built This
I was using Claude Code with Drizzle ORM and hit a wall: drizzle-kit generate asks interactive questions like "Is this table created or renamed?" that Claude Code couldn't answer. It just hung. I wanted something minimal and self-contained, so I built one.
It works for way more than just Drizzle though — any CLI tool with interactive prompts becomes fully autonomous:
- Database migrations — Drizzle Kit, Prisma, TypeORM, Knex
- Project scaffolding —
npm init,create-next-app,create-vite,npx degit - Package managers —
npm installpeer dep prompts,yarnresolutions - Git operations — interactive rebase, merge conflict resolution,
git add -p - Cloud CLIs —
aws configure,gcloud init,firebase init,vercel - Docker —
docker buildprompts,docker composeconfirmations - System tools —
ssh-keygen,gpg --gen-key,certbot - Linters/formatters — ESLint
--init, Prettier setup,stylelintconfig
The Problem
AI coding agents like Claude Code can run shell commands, but they can't handle interactive prompts. When a CLI tool asks "Is this table created or renamed?" or "Pick a preset:", the agent gets stuck — it can't read the prompt or type an answer. This blocks any CLI workflow that requires human input — database migrations, project scaffolding, package configuration, and more.
The Solution
pty-mcp gives the agent a real pseudo-terminal via the Model Context Protocol. The agent can:
- Spawn a command in a PTY
- Read the interactive prompt output
- Write an intelligent response
- Repeat until the process exits
Setup
Add to your Claude Code MCP settings (~/.claude/settings.json):
{
"mcpServers": {
"pty-mcp": {
"command": "npx",
"args": ["-y", "pty-mcp"]
}
}
}
Requirements
- Node.js 20+
- Build tools for native addon compilation (Python 3, make, g++)
- macOS:
xcode-select --install - Ubuntu/Debian:
sudo apt install build-essential python3 - Windows: Pre-built binaries included, no extra tools needed
- macOS:
Tools
pty_spawn
Spawn a command in a pseudo-terminal.
| Parameter | Type | Default | Description |
|---|---|---|---|
command |
string | (required) | Command to run |
args |
string[] | [] |
Command arguments |
cwd |
string | server CWD | Working directory |
env_vars |
object | {} |
Extra environment variables (merged with system env) |
idle_timeout_ms |
number | 3000 |
Wait for output to settle before returning |
Returns: { session_id, output, is_running, exit_code }
pty_write
Send input to a running PTY session.
| Parameter | Type | Default | Description |
|---|---|---|---|
session_id |
string | (required) | Session ID from pty_spawn |
input |
string | (required) | Text to send |
press_enter |
boolean | true |
Append Enter after input |
idle_timeout_ms |
number | 2000 |
Wait for output to settle before returning |
Returns: { output, is_running, exit_code }
pty_kill
Kill a running session.
| Parameter | Type | Default | Description |
|---|---|---|---|
session_id |
string | (required) | Session ID to kill |
Returns: { success: true }
Example: Handling Interactive Prompts
Agent calls pty_spawn({ command: "npx", args: ["drizzle-kit", "generate"] })
→ Returns output: "Is 'users' table created or renamed from another table? ❯ create / rename"
Agent reads the prompt, understands context, decides "create"
→ Calls pty_write({ session_id: "abc-123", input: "" })
Process continues, agent answers more prompts as needed...
Process exits → agent gets final output with results
How It Works
- Uses node-pty (Microsoft, powers VS Code's terminal) for real PTY allocation
- ANSI escape codes are stripped automatically for clean output
- Sessions auto-expire after 5 minutes of inactivity
- All sessions are cleaned up on server shutdown
- No shell wrapping — commands are spawned directly (no injection risk)
Security
- Local only — stdio transport, no network exposure, no ports opened
- No shell injection — uses
pty.spawn(command, args)directly, notbash -c - No secrets stored — environment variables are passed through, not logged
- Session isolation — each spawn gets its own PTY with a unique session ID
- Session limits — max 20 concurrent sessions, 30s max idle timeout per request
- Auto-cleanup — idle sessions killed after 5 minutes, graceful shutdown on crash
Trust model: This server grants command execution to the connected MCP client. Only connect it to clients you trust (e.g., Claude Code on your local machine). The server inherits your shell environment — spawned processes have access to the same env vars as your terminal.
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。