MCP Python AI Agents

How to Turn Any Python CLI
into an MCP Server

May 2026 · 8 min read · click-to-mcp on GitHub

MCP (Model Context Protocol) is how AI coding agents talk to tools. Claude Code, Cursor, Codex, Cline — they all use it. But if your team already has Python CLI tools built with Click or typer, you'd normally need to write a separate MCP server for each one.

That means duplicate code, duplicate maintenance, and a new server for every tool. Not sustainable.

Here's how to skip all of that and turn your existing Python CLI into an MCP server with zero code changes.

The Three Ways to Make an MCP Server

Approach	Effort	Best For
Write from scratch (FastMCP)	Medium	New tools, greenfield
Manual wrapper per CLI	High	1-2 CLIs with custom logic
Auto-wrap with click-to-mcp	Low	Existing Click/typer CLIs

If you're building something new, FastMCP is great. But most teams have existing tools. That's where click-to-mcp shines — it introspects your CLI at runtime and serves it as an MCP server automatically.

Step 1: Install click-to-mcp

pip install git+https://github.com/Coding-Dev-Tools/click-to-mcp.git

If you want HTTP+SSE transport (for web-based clients like Cursor), install with the http extra:

pip install "click-to-mcp[http] @ git+https://github.com/Coding-Dev-Tools/click-to-mcp.git"

Step 2: Discover Your CLIs

$ click-to-mcp discover
Scanning environment for Click/typer CLIs...

Found 3 CLIs:
  • my-database-tool (6 commands)
  • my-api-validator (4 commands)
  • my-deploy-helper (3 commands)

click-to-mcp scans your Python environment and finds every installed Click or typer CLI. No configuration needed — it detects them automatically.

Step 3: Serve as MCP Server

$ click-to-mcp serve my-database-tool
MCP server starting on stdio...
  Tool: db_query (2 params)
  Tool: db_migrate (1 param)
  Tool: db_seed (1 param)
  Tool: db_dump (1 param)
  Tool: db_stats (0 params)
  Tool: db_reset (0 params)
Ready — 6 tools exposed

Every command in your CLI becomes an MCP tool. Arguments become required input properties. Options become optional properties with defaults. click.Choice becomes a JSON Schema enum. Nested groups become prefixed tools.

How Click Maps to MCP

Click Concept	MCP Mapping
`@click.command()`	MCP tool
`@click.argument()`	Required input property
`@click.option()`	Optional input with default
`click.Choice`	JSON Schema `enum`
`click.INT` / `click.FLOAT`	`integer` / `number`
`click.BOOL` / `is_flag`	`boolean`
Nested `click.Group`	Prefixed tools (e.g. `config_show`)

The mapping is automatic because Click's type system already carries everything MCP needs. No annotations. No decorators. No boilerplate.

Step 4: Connect to Your AI Agent

Claude Code

Add to .claude/settings.json:

{
  "mcpServers": {
    "my-database-tool": {
      "command": "click-to-mcp",
      "args": ["serve", "my-database-tool"]
    }
  }
}

Cursor

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "my-database-tool": {
      "command": "click-to-mcp",
      "args": ["serve", "my-database-tool"]
    }
  }
}

Cline / VS Code

Add to .vscode/mcp.json:

{
  "servers": {
    "my-database-tool": {
      "command": "click-to-mcp",
      "args": ["serve", "my-database-tool"]
    }
  }
}

Step 5: Use It

Now when you ask your AI agent a question about your tool, it will automatically invoke the MCP tool:

You: What tables are in my database?

Claude Code: *calls db_stats tool*
→ Database: production_db
→ 23 tables, 1.2M rows total
→ Largest: orders (450K rows), users (180K rows)

No manual command invocation. The AI agent picks the right tool and arguments automatically.

stdio vs HTTP+SSE: Which Transport?

Transport	Command	Use With
stdio (default)	`click-to-mcp serve my-cli`	Claude Code, Cline, terminal agents
HTTP+SSE	`click-to-mcp serve-http my-cli --port 8000`	Cursor, web clients, remote agents

Most local development uses stdio. If your AI tool connects over HTTP (like Cursor does for some configurations), use the HTTP+SSE transport.

Multiple CLIs at Once

The real power is exposing your entire toolkit:

{
  "mcpServers": {
    "database-tool": {
      "command": "click-to-mcp",
      "args": ["serve", "my-database-tool"]
    },
    "api-validator": {
      "command": "click-to-mcp",
      "args": ["serve", "my-api-validator"]
    },
    "deploy-helper": {
      "command": "click-to-mcp",
      "args": ["serve", "my-deploy-helper"]
    }
  }
}

Now your AI agent has access to all three tools — database operations, API validation, and deployment commands — in a single conversation.

What click-to-mcp Is NOT

Not a replacement for FastMCP. Building a new MCP server from scratch? Use FastMCP. click-to-mcp is for existing CLIs.
Not limited to our tools. Any Click/typer CLI works — your internal tools, open-source packages, anything pip-installable.
Not a runtime proxy. It introspects at startup, then serves directly. Zero per-request overhead.
Not a code generator. Your CLI code stays untouched. click-to-mcp reads the Click registry at runtime.

Who Should Use This

Teams with multiple Python CLIs — wrap them all as MCP servers in minutes, not days
DevOps teams — expose deployment, monitoring, and infrastructure CLIs to AI agents
API developers — let AI agents validate specs, run checks, and diff contracts
Anyone using AI coding agents — give your agent access to your actual tools, not just file editing

Try it now — your CLI becomes an MCP server in 10 seconds

Star on GitHub →

How to Turn Any Python CLIinto an MCP Server