> For the complete documentation index, see [llms.txt](https://developer.paddle.com/llms.txt).

# The Paddle MCP server is now remote

The Paddle MCP server is now a hosted, remote service. No local install, faster setup, and a new codemode interface that gives agents access to the full Paddle API through three tools.

---

## What's new?

The [Paddle MCP server](https://developer.paddle.com/sdks/ai/paddle-mcp.md) is now a remote, hosted service. Your agent connects over HTTPS using your Paddle API key as a Bearer token.

We host one endpoint per environment:

| Environment | URL                                  |
| ----------- | ------------------------------------ |
| Sandbox     | `https://sandbox-mcp.paddle.com/mcp` |
| Live        | `https://mcp.paddle.com/mcp`         |

Both work today across [Claude Code](https://developer.paddle.com/get-started/ai/claude-code.md), Claude Desktop, [Cursor](https://developer.paddle.com/get-started/ai/cursor.md), VS Code, Codex, Mistral Vibe, and any other MCP-compatible client.

## How it works

Previously, we released [`@PaddleHQ/paddle-mcp-server`](PaddleHQ/paddle-mcp-server), an MCP server that ran locally on your device.

Today, we've released our remote MCP server. Instead of running on your machine, it's hosted on our servers. This means that it's always up-to-date with the latest Paddle API operations, and unlocks agentic workflows. For example, you can call the MCP from a CI job, [Claude Code routine](https://code.claude.com/docs/en/routines), or any agentic workflow that runs in a remote environment.

The new server uses [a codemode interface](https://blog.cloudflare.com/code-mode/). Instead of exposing every API operation as a separate MCP tool, it exposes three:

- `search`: look up Paddle API methods by name or description.
- `execute`: run a JavaScript async function that chains one or more Paddle API calls in a single invocation.
- `report_missing_tool`: flag a missing capability for us to prioritize.

Codemode is more efficient because it allows the agent to write code to call the tools, rather than calling the tools directly. [Research shows](https://github.com/imran31415/codemode_python_benchmark) that codemode results in 60% faster execution, 68% fewer tokens, and 88% fewer API round trips while maintaining equal accuracy.

You don't need to do anything differently to use the remote server. The server handles the codemode interface for you.

## Migrate from the local server

If you're running `npx @paddle/paddle-mcp-server` locally, we recommend migrating to the remote server when you get a chance. The local `@paddle/paddle-mcp-server` npm package still works today, but we won't be updating it with new or changed Paddle API operations. We'll deprecate it in a future release.

Update your MCP client config with the snippet for your editor of choice from [the install instructions](https://developer.paddle.com/sdks/ai/paddle-mcp#install-the-mcp-server.md). We recommend wiring both `paddle-sandbox` and `paddle-live` under separate server names so the agent can route prompts to the right environment.

The remote MCP doesn't support tool call filtering. If you need to restrict the tools that the agent can call, you can scope your API key's permissions.

## Next steps

The remote Paddle MCP server is available now. You can use it with any MCP-compatible client, including Claude Code, Claude Desktop, Cursor, VS Code, Codex, Raycast, and any other MCP-compatible client.

If you use Claude Code, you can install the Paddle agent skills plugin to get the remote server and all the agent skills in one go. For details, see [Build with Claude Code](https://developer.paddle.com/get-started/ai/claude-code.md).

See [Paddle MCP server](https://developer.paddle.com/sdks/ai/paddle-mcp.md) for install instructions, details about codemode, and example prompts.