---

name: executor description: "Manage persistent interactive CLI processes (REPLs, debuggers, database CLIs, NuttX simulator, QEMU). Use when process maintains state across commands. NOT for one-off commands - use Bash. Prefer over tmux for programmatic I/O without terminal emulation." license: Apache-2.0 compatibility: "Requires executor-mcp MCP server. Install: pip install executor-mcp or from source."

Executor

Manage persistent interactive CLI processes with stateful stdin/stdout communication via MCP tools.

Pattern: Start → Send* → Read* → Stop

Prerequisites

This skill requires the executor-mcp MCP server to be installed and configured.

Installation:

Copypip install executor-mcp

Configuration: Add to Claude Desktop config (~/.config/Claude/claude_desktop_config.json):

Copy{
  "mcpServers": {
    "executor": {
      "command": "executor-mcp"
    }
  }
}

When to Use

Use executor when:

• Process has a prompt and maintains state (Python REPL, gdb, psql, node)
• You need programmatic access to stdin/stdout without terminal complexity
• You want automatic buffering and logging of all I/O

Core Tools

All tools return JSON responses. Parameters shown below.

executor_start

Launch a new interactive process.

Parameters:

Copy{
  "command": "python3",
  "args": ["-i"],
  "working_dir": "/optional/path"
}

Returns: process_id (unique identifier for subsequent operations)

executor_send

Send text to process stdin and optionally wait for output.

Parameters:

Copy{
  "process_id": "abc123",
  "text": "x = 42",
  "wait_time": 0.1,
  "add_newline": true,
  "tail_lines": 20,
  "full_buffer": false
}

Key parameters:

• wait_time: Seconds to wait before reading (default: 0.1)
  • > 0: Wait and return NEW output from this command
  • 0: Send immediately, use executor_read_output later
• full_buffer: If true, return full buffer instead of just new output (default: false)
• tail_lines: Lines to return when using full_buffer=true (default: 20)

Returns: New output lines if wait_time > 0, otherwise "Success"

executor_read_output

Read buffered output from stdout/stderr.

Parameters:

Copy{
  "process_id": "abc123",
  "tail_lines": 50,
  "stream": "both"
}

Parameters:

• tail_lines: Number of recent lines (default: all buffered, max 1000)
• stream: "both" (merged), "stdout", or "stderr" (default: "both")

Returns: Buffered output lines (last 1000 lines kept in memory)

executor_stop

Terminate a running process.

Parameters:

Copy{
  "process_id": "abc123",
  "force": false
}

• force: true = SIGKILL, false = SIGTERM graceful (default)

executor_list

List all active processes (no parameters).

Returns: Array of process info (process_id, command, status, buffer sizes)

executor_get_info

Get detailed information about a specific process.

Parameters:

Copy{
  "process_id": "abc123"
}

Returns: Full process state, buffer sizes, recent output preview

Examples

Python REPL

Copy# Start Python with unbuffered output
executor_start(command="python3", args=["-i", "-u"])
# → Returns: {"process_id": "abc123", ...}

# Wait for startup banner (0.3s)
executor_send(process_id="abc123", text="x = 42", wait_time=0.3)
# → Returns new output: [">>> ", "x = 42\n", ">>> "]

# Execute and read output
executor_send(process_id="abc123", text="print(x * 2)")
# → Returns: ["print(x * 2)\n", "84\n", ">>> "]

# Stop when done
executor_stop(process_id="abc123")

GDB Debugging

Copy# Start debugger
executor_start(command="gdb", args=["--quiet", "./binary"])
# → process_id: "def456"

# Disable pagination
executor_send(process_id="def456", text="set pagination off", wait_time=0.2)

# Set breakpoint
executor_send(process_id="def456", text="break main", wait_time=0.2)

# Run program
executor_send(process_id="def456", text="run", wait_time=0.5)

# Examine variables
executor_send(process_id="def456", text="info locals")

# Exit debugger
executor_stop(process_id="def456")

Database CLI (PostgreSQL)

Copy# Connect to database
executor_start(command="psql", args=["dbname", "-U", "user"])
# → process_id: "ghi789"

# Run query (wait for results)
executor_send(
    process_id="ghi789",
    text="SELECT * FROM users LIMIT 5;",
    wait_time=0.5
)
# → Returns query results

# List tables
executor_send(process_id="ghi789", text="\\dt")

# Disconnect
executor_send(process_id="ghi789", text="\\q", wait_time=0)
executor_stop(process_id="ghi789")

Fast Batch Commands (No Wait)

Copy# Start process
executor_start(command="python3", args=["-i"])
# → process_id: "jkl012"

# Queue multiple commands without waiting
executor_send(process_id="jkl012", text="import sys", wait_time=0)
executor_send(process_id="jkl012", text="import os", wait_time=0)
executor_send(process_id="jkl012", text="import json", wait_time=0)

# Read all output at once
output = executor_read_output(process_id="jkl012", tail_lines=50)

NuttX Simulator

Copy# Start NuttX simulator
executor_start(command="./nuttx/nuttx", working_dir="/path/to/nuttx")
# → process_id: "mno345"

# Wait for NSH prompt
executor_send(process_id="mno345", text="help", wait_time=0.5)

# Run NuttX commands
executor_send(process_id="mno345", text="ps")
executor_send(process_id="mno345", text="free")

# Exit simulator
executor_send(process_id="mno345", text="exit", wait_time=0)
executor_stop(process_id="mno345", force=true)

Complete Workflow Example

Debugging a Python script that crashes:

Copy# 1. Start Python debugger with script
executor_start(command="python3", args=["-i", "buggy_script.py"])
# → process_id: "debug001"

# 2. Script crashes, now inspect
executor_send(process_id="debug001", text="import traceback")
executor_send(process_id="debug001", text="traceback.print_exc()")

# 3. Inspect variables at crash point
executor_send(process_id="debug001", text="print(locals())")

# 4. Test fix interactively
executor_send(process_id="debug001", text="x = corrected_value")
executor_send(process_id="debug001", text="retry_operation(x)")

# 5. Verify fix works
output = executor_read_output(process_id="debug001", tail_lines=20)

# 6. Exit when done
executor_stop(process_id="debug001")

Error Handling

Common errors and solutions:

Debugging steps:

1. Check if process is still running: executor_get_info(process_id)
2. Review complete I/O history in log files: .executorlog/{process_id}_{timestamp}_{command}.log
3. Verify process exit code and error stream
4. For slow processes, increase wait_time (0.3-0.5s)

Best Practices

• Always use unbuffered mode for Python: python3 -u or python3 -i
• Adjust wait_time based on command complexity:
  • Fast commands (variable assignment): 0.1s (default)
  • Medium commands (queries, calculations): 0.3s
  • Slow commands (compilations, heavy queries): 0.5s+
• Use wait_time=0 for batch operations when you don't need immediate output
• Always call executor_stop when done to prevent orphaned processes
• Check logs (.executorlog/) for debugging - all I/O is timestamped
• Use executor_list periodically to audit running processes

Logging

All I/O is automatically logged to:

Location: .executorlog/{process_id}_{timestamp}_{command}.log

Format:

Copy=== Executor MCP Process Log ===
Process ID: abc123
Command: python3
Started: 2026-01-07T12:00:00
==================================================

[2026-01-07 12:00:00.123] COMMAND: python3 -i -u
[2026-01-07 12:00:00.456] STDOUT: Python 3.13.0 ...
[2026-01-07 12:00:01.789] STDIN: x = 42
[2026-01-07 12:00:01.890] STDOUT: >>>
[2026-01-07 12:00:02.012] STDIN: print(x * 2)
[2026-01-07 12:00:02.123] STDOUT: 84
[2026-01-07 12:00:05.567] TERMINATED: Method: SIGTERM, Return code: 0

Configure log directory:

Copyexport EXECUTOR_LOG_DIR="$HOME/.executor-mcp/logs"

Architecture Notes

• Buffer: Last 1000 lines kept in memory per stream (stdout/stderr)
• Concurrency: Multiple processes can run simultaneously
• Non-blocking: All operations use async I/O
• Transport: MCP stdio (separate from managed process I/O)
• Logging: Complete history in files, circular buffer in memory