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 install peer dep prompts, yarn resolutions
• Git operations — interactive rebase, merge conflict resolution, git add -p
• Cloud CLIs — aws configure, gcloud init, firebase init, vercel
• Docker — docker build prompts, docker compose confirmations
• System tools — ssh-keygen, gpg --gen-key, certbot
• Linters/formatters — ESLint --init, Prettier setup, stylelint config

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:

1. Spawn a command in a PTY
2. Read the interactive prompt output
3. Write an intelligent response
4. 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

Tools

pty_spawn

Spawn a command in a pseudo-terminal.

Returns: { session_id, output, is_running, exit_code }

pty_write

Send input to a running PTY session.

Returns: { output, is_running, exit_code }

pty_kill

Kill a running session.

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, not bash -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