Claude Code MCP Server

An MCP (Model Context Protocol) server that allows running Claude Code in one-shot mode with permissions bypassed automatically.

Screenshot

Overview

Cursor Screenshot

Did you notice that Cursor often struggles to apply smaller edits via edit_file, especially when linting? And then it tries multiple times till it eventually works? Yeah... no more. It LOVES magic_file.

This MCP server provides two tools that can be used by LLMs to interact with Claude Code. When integrated with Claude Desktop or other MCP clients, it allows LLMs to:

Run Claude Code with all permissions bypassed (using --dangerously-skip-permissions)
Execute Claude Code with any prompt without permission interruptions
Access file editing capabilities directly
Enable specific tools by default

Prerequisites

Node.js v16 or later
TypeScript (for development)
Claude CLI installed and working. Ensure Claude CLI is installed and accessible, preferably by running /doctor. This installs/updates the CLI to ~/.claude/local/claude, which this server checks by default.

Installation

Clone or download this repository:

git clone https://github.com/yourusername/claude-mcp-server.git
cd claude-mcp-server

Install dependencies (this will also install tsx for direct TypeScript execution):

npm install

Make the start script executable:

chmod +x start.sh

Connecting to Cursor/Windsurf/Visual Studio Code

macOS

Locate the MCP configuration file:
```
~/.cursor/mcp.json
```
Create this file if it doesn't exist.
Add the server configuration to the file:

{
  "mcpServers": {
    "claude_code": {
      "type": "stdio",
      "command": "/absolute/path/to/claude-mcp-server/start.sh",
      "args": []
    }
  }
}

Make sure to replace /absolute/path/to/claude-mcp-server with the actual path to where you installed this server.

Windows

Locate the MCP configuration file:
```
%APPDATA%\cursor\mcp.json
```
Create this file if it doesn't exist.
Add the server configuration to the file, making sure to use Windows path format:

{
  "mcpServers": {
    "claude_code": {
      "type": "stdio",
      "command": "C:\\path\\to\\claude-mcp-server\\start.bat",
      "args": []
    }
  }
}

For Windows, you should use the batch file (start.bat) instead of the shell script.

Linux

Locate the MCP configuration file:
```
~/.config/cursor/mcp.json
```
Create this file if it doesn't exist.
Add the server configuration to the file:

{
  "mcpServers": {
    "claude_code": {
      "type": "stdio",
      "command": "/absolute/path/to/claude-mcp-server/start.sh",
      "args": []
    }
  }
}

After updating the configuration, restart your IDE to load the MCP server.

Environment Variables

You can customize the server behavior with the following environment variables (edit them in start.sh or start.bat):

CLAUDE_CLI_PATH: Optional. Set a custom absolute path to the Claude CLI executable. If set and the path points to an existing file, this path will be used directly.
MCP_CLAUDE_DEBUG: Set to true to enable verbose debug logging from the server to stderr (e.g., MCP_CLAUDE_DEBUG=true ./start.sh).

Claude CLI Discovery Order:

The path specified by the CLAUDE_CLI_PATH environment variable (if set and valid).
The default installation path for Unix-like systems: ~/.claude/local/claude (where ~ is the user's home directory). For Windows users, this automatic check may not apply; relying on CLAUDE_CLI_PATH or ensuring claude is in the system PATH is recommended.
Defaults to simply claude, relying on the system's PATH for resolution (a warning will be logged if this fallback is used).

Connecting to VSCode Claude

To use this MCP server with Claude in VSCode:

Install the Claude extension in VSCode

Create or edit the MCP settings file:

~/.vscode/extensions/saoudrizwan.claude-dev/settings/cline_mcp_settings.json

Add the server configuration:

{
  "mcpServers": {
    "claude_code": {
      "command": "/absolute/path/to/claude-mcp-server/start.sh",
      "args": [],
      "disabled": false
    }
  }
}

Usage

Once installed and connected to an MCP client, you can invoke the tools using the following formats:

Claude Code Tool (renamed to `code`)

{
  "tool_name": "code",
  "params": {
    "prompt": "Your prompt to Claude Code here",
    "options": {
      "tools": ["Bash", "Read", "Write"]
    }
  }
}

If no tools are specified, the server enables common tools by default.

Claude File Edit Tool (renamed to `magic_file`)

{
  "tool_name": "magic_file",
  "params": {
    "file_path": "/path/to/your/file.js",
    "instruction": "Refactor the processData function to use async/await instead of promises."
  }
}

Tool Descriptions

The server provides two tools:

Tool name: code
- Description: "Executes a given prompt directly with the Claude Code CLI, bypassing all permission checks (--dangerously-skip-permissions). Ideal for a wide range of tasks including: complex code generation, analysis, and refactoring; performing web searches and summarizing content; running arbitrary terminal commands (e.g., open . to open Finder, open -a Calculator to open apps, or open https://example.com to open a URL in a web browser). For example, you could open a GitHub PR page once all tests are green. Handles general tasks requiring the Claude CLI's broad capabilities without interactive prompts. options.tools can be used to specify internal Claude tools (e.g., Bash, Read, Write); common tools are enabled by default if this is omitted."
- Parameters:
  - prompt (required): The prompt to send to Claude Code
  - options.tools (optional): Array of specific tools to enable
- Implementation: Uses claude --dangerously-skip-permissions (invoked via child_process.spawn) to bypass all permission checks. The server locates the Claude CLI by first checking the CLAUDE_CLI_PATH environment variable, then looking in ~/.claude/local/claude, and finally falling back to claude in the system PATH.
Tool name: magic_file
- Description: "Edits a specified file based on natural language instructions, leveraging the Claude Code CLI with all editing permissions bypassed (--dangerously-skip-permissions). Best for complex or semantic file modifications where describing the desired change in plain language is more effective than precise line-by-line edits. Requires an absolute file_path and a descriptive instruction. Also a great alternative if a general-purpose edit_file tool is struggling with complex edits or specific file types. Example instruction: 'Refactor the processData function to use async/await instead of promises.'"
- Parameters:
  - file_path (required): The absolute path to the file to edit
  - instruction (required): Free text description of the edits to make to the file
- Implementation: Uses claude --dangerously-skip-permissions (invoked via child_process.spawn) with Edit tools enabled. The server locates the Claude CLI by first checking the CLAUDE_CLI_PATH environment variable, then looking in ~/.claude/local/claude, and finally falling back to claude in the system PATH.

Troubleshooting

Tool not showing up: Check the Claude logs for errors when starting the MCP server. Ensure start.sh or start.bat is executable and tsx is installed and runnable (usually via npx).
Command not found / "Error: spawn claude ENOENT" / "[Warning] Claude CLI not found... Falling back to "claude" in PATH...": This means the server could not find the claude executable via the CLAUDE_CLI_PATH environment variable (if set), at the default Unix-like location (~/.claude/local/claude), or in the system PATH (if it fell back to just 'claude').
- Ensure the Claude CLI is installed correctly. For Unix-like systems, this is often at ~/.claude/local/claude (verify by running /doctor in a Claude context). For Windows, ensure it's in your system PATH or set CLAUDE_CLI_PATH.
- Explicitly set the CLAUDE_CLI_PATH environment variable in start.sh or start.bat to the correct absolute path of your claude executable.
Permission errors: Ensure the start.sh script is executable and that Node.js has permission to execute tsx and the Claude CLI (whether found via CLAUDE_CLI_PATH, the default path, or the system PATH).

License

MIT

Server test complete.

Claude Code MCP Server