MCP server that gives AI coding agents (Claude Code, Cursor, etc.) the ability to interact with interactive CLI processes via PTY — answering prompts, navigating menus, and controlling TUI applications.
Strong tool definitions with comprehensive descriptions (avg 180 chars) and well-structured Zod schemas. All 8 tools have clear verb-noun naming (spawn, send_input, send_keys, get_screen, wait_for, resize, close, list_sessions). Tool annotations present (destructiveHint, readOnlyHint, openWorldHint). However, output schemas are not formally documented in the code, responses are JSON-stringified but lack explicit schema declarations. Error handling is basic (throws Error with context but no recovery guidance). Parameter descriptions are strong but some lack explicit constraints (e.g., waitMs lacks min/max bounds). No pagination support for list_sessions despite potential for many active sessions.
Terminate a session and clean up resources. Sends SIGTERM to the process; if it doesn't exit within 2 seconds, sends SIGKILL. Always succeeds.
Get the current screen output from a running process. Optionally search for text patterns to locate specific content (e.g. find error messages, prompts, status lines). Returns the full screen and match positions.
List all active sessions with metadata (command, uptime, bytes received, etc.). Useful for debugging or monitoring long-running processes.
Resize the terminal for a running session. Useful if the process needs to adapt to different screen dimensions (e.g. full-screen TUI apps).
Type text into the interactive process and return the updated screen. Appends Enter (newline) automatically. Set pressEnter=false to type without submitting — useful for password fields or filling forms before tabbing.
Send one or more special keys in sequence to navigate interactive TUI menus, confirm dialogs, or control the process. A small delay between keys ensures TUI frameworks register each keystroke. Examples: ["down","down","enter"] — select 3rd item in a menu ["tab","tab","enter"] — skip to submit button ["ctrl+c"] — cancel/interrupt the process ["y","enter"] — confirm a yes/no prompt
Output schemas not formally documented. Tool responses are JSON-stringified but lack explicit schema declarations in tool registration. LLMs cannot reliably parse response structure without documented schemas.
Numeric parameters lack explicit bounds. waitMs, cols, rows, delayBetweenMs, timeoutMs have no min/max constraints documented. LLMs may pass absurd values (negative, zero, or millions of ms).
Error handling lacks recovery guidance. Errors throw generic Error objects with context but do not categorize as retryable/user-fixable/fatal or suggest next steps. E.g., 'Session not found' should suggest 'Call list_sessions() to see active sessions.'
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | 2026-07-28+ | v2 |
Start an interactive CLI process in a real PTY with virtual terminal rendering. Returns a session ID and the initial screen output. Use for any command that requires user interaction: eas build, gcloud auth, npm init, ssh, docker login, interactive installers, REPLs (python, node, psql), etc. The "screen" field shows exactly what a human would see — clean text with TUI menus, prompts, progress bars, and selections properly rendered.
Wait for specific text to appear on the screen (blocking). Useful for waiting for prompts, status messages, or completion indicators before proceeding. Returns immediately if text is already visible.
list_sessions lacks pagination. No limit, offset, or cursor parameters. If many sessions accumulate, response could exceed context window. Should support limit (default 20) and offset.
No confirmation/dry-run for destructive operations. spawn and close are marked destructive but lack a confirm_before_execute pattern. Agents could accidentally kill critical processes.