MCP Ask-to-Code

An autonomous MCP server that uses the Claude Code CLI to solve coding problems without permission prompts.

Quick Install

# Global install with pipx (recommended)
pipx install git+https://github.com/cjaviersaldana/mcp-ask-to-code.git

# Or with pip
pip install git+https://github.com/cjaviersaldana/mcp-ask-to-code.git

Prerequisites

Claude Code CLI installed and authenticated:

npm install -g @anthropic-ai/claude-code

Usage

# Run (stdio mode for Claude Desktop)
mcp-ask-to-code

# Run with a specific model
mcp-ask-to-code --model claude-sonnet-4-20250514

# Run on a port (SSE mode for multiple sessions)
mcp-ask-to-code --port 8080

Options

Argument	Env Variable	Default	Description
--model	CLAUDE_MODEL	(CLI default)	Model to use (e.g. claude-sonnet-4-20250514, opus)
--name	MCP_NAME	Autonomous Code Agent	Server name
--tool-name	MCP_TOOL_NAME	ask_autonomous_agent	Tool function name
--command	CLAUDE_CMD	claude	Claude CLI binary path
--port	MCP_PORT	0 (stdio)	Port for SSE mode
--host	MCP_HOST	127.0.0.1	Host for SSE mode

Claude Desktop Configuration

Add to your config file:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Linux: ~/.config/claude-desktop/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

Basic Setup

{
  "mcpServers": {
    "ask-to-code": {
      "command": "mcp-ask-to-code",
      "args": []
    }
  }
}

With Specific Model

{
  "mcpServers": {
    "ask-to-code": {
      "command": "mcp-ask-to-code",
      "args": ["--model", "claude-sonnet-4-20250514"]
    }
  }
}

Custom "Expert" Personality

{
  "mcpServers": {
    "backend-expert": {
      "command": "mcp-ask-to-code",
      "args": ["--model", "opus", "--name", "Backend Architect"],
      "env": {
        "MCP_TOOL_NAME": "consult_backend_expert"
      }
    }
  }
}

Multiple Agents

{
  "mcpServers": {
    "code-agent": {
      "command": "mcp-ask-to-code",
      "args": ["--model", "claude-sonnet-4-20250514"]
    },
    "code-reviewer": {
      "command": "mcp-ask-to-code",
      "args": ["--model", "opus"],
      "env": {
        "MCP_NAME": "Code Reviewer",
        "MCP_TOOL_NAME": "review_code"
      }
    }
  }
}

How It Works

1. Exposes a single tool (ask_autonomous_agent) to the LLM
2. Delegates questions to Claude Code CLI with --dangerously-skip-permissions -p
3. Returns clean output (ANSI codes stripped)

License

MIT