cli2mcp -- Turn any CLI into an MCP Server

An educational Python library that bridges the gap between traditional command-line tools and the Model Context Protocol (MCP).

cli2mcp scans a CLI tool's --help output, extracts its arguments and subcommands, and generates a JSON descriptor file. It can then serve that file as a fully functional MCP server -- letting AI assistants call the CLI tool through a standard protocol.

Quick start

uv sync

1. Scan a CLI tool

uv run cli2mcp scan curl

This runs curl --help, parses the output, and writes curl.tools.json.

For tools with subcommands (like git):

uv run cli2mcp scan git -o git.tools.json

Each subcommand becomes its own MCP tool.

2. Serve as an MCP server

uv run cli2mcp serve curl.tools.json

This starts an MCP server (stdio transport) that exposes every entry in the JSON file as a callable tool.

To use HTTP transport instead:

uv run cli2mcp serve curl.tools.json -t streamable-http

3. Connect to an AI assistant

Add the server to your assistant's MCP configuration.

{
  "mcpServers": {
    "curl": {
      "command": "cli2mcp",
      "args": ["serve", "curl.tools.json"]
    }
  }
}

Related MCP server: mcp-cli-catalog

How it works

                                        MCP client
                                            |
cli2mcp scan curl   -->  curl.tools.json    |
                              |             |
                         cli2mcp serve  <---+
                              |
                         subprocess.run(["curl", ...])

The JSON schema

The generated file looks like this:

{
  "command": "curl",
  "tools": [
    {
      "name": "curl",
      "description": "transfer a URL",
      "args": [
        {
          "name": "url",
          "description": "URL to transfer",
          "type": "string",
          "required": true
        },
        {
          "name": "--output",
          "description": "Write output to file instead of stdout",
          "type": "string",
          "required": false
        }
      ]
    }
  ]
}

• command -- the base CLI binary to run.

• tools -- one entry per tool (or per subcommand).

• args -- each argument has a name, description, type (always "string"), and required flag.

• Argument names starting with -- are flags; others are positional.

• You can hand-edit this file to add, remove, or rename tools.

How arguments map back to CLI commands

When the MCP server receives a tool call like:

{"name": "git_commit", "arguments": {"message": "fix bug", "all": "true"}}

It reconstructs the CLI command:

git commit --message "fix bug" --all true

Flags (names starting with --) are emitted as --flag value. Positional arguments are appended at the end.

Supported help styles

Different CLI frameworks produce different --help formats. cli2mcp auto-detects the style and uses the right parser:

Requirements

• Python 3.10+

• uv (recommended) or pip

• mcp[cli] (the official MCP Python SDK, installed automatically)