> ## Documentation Index
> Fetch the complete documentation index at: https://docs.creem.io/llms.txt
> Use this file to discover all available pages before exploring further.

# MCP Server

> Give your AI agent 49 tools that call the Creem API directly, over the Model Context Protocol.

The Creem TypeScript SDK ships an [MCP](https://modelcontextprotocol.io) server. Point Claude Code, Claude Desktop, Cursor or any MCP client at it and your agent can create products, open checkouts, manage subscriptions, issue refunds and inspect your store — by calling the real API, with your key.

<Tip>
  **Skill files vs MCP server.** [Skill files](/ai/for-agents/skill-files) are reference material your agent *reads* — they teach it how Creem works. The MCP server is a set of tools your agent *runs* — it changes live data. Most teams want both.
</Tip>

## Requirements

* **Node.js v22 or newer**
* A Creem API key ([test or live](/getting-started/test-mode))

Nothing to install: `npx` fetches the `creem` package on demand.

## Quickstart

```bash theme={null}
npx -y --package creem -- mcp start --api-key creem_test_your_key --server test
```

The server speaks MCP over stdio and exits when the client disconnects.

<Warning>
  **Test-mode keys require `--server test`.** Without it the SDK targets production, and every tool call fails with `401 Unauthorized — Invalid API Key`, even though your key is valid. This is the single most common setup problem.
</Warning>

## Claude Code

```bash theme={null}
claude mcp add creem -- npx -y --package creem -- mcp start --api-key creem_test_your_key --server test
```

## Claude Desktop

Add to `claude_desktop_config.json`:

```json theme={null}
{
  "mcpServers": {
    "Creem": {
      "command": "npx",
      "args": [
        "-y", "--package", "creem", "--",
        "mcp", "start",
        "--api-key", "creem_test_your_key",
        "--server", "test"
      ]
    }
  }
}
```

## Cursor

Add the same block to `.cursor/mcp.json` in your project, or to `~/.cursor/mcp.json` for every project.

## Available tools

49 tools, covering the API surface.

| Group                | Count | Tools                                                                                                                                                                                                                               |
| -------------------- | ----- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Products**         | 5     | `products-create`, `products-get`, `products-update`, `products-search`, `products-archive`                                                                                                                                         |
| **Checkouts**        | 2     | `checkouts-create`, `checkouts-retrieve`                                                                                                                                                                                            |
| **Customers**        | 8     | `customers-create`, `customers-retrieve`, `customers-update`, `customers-list`, `customers-get-orders`, `customers-list-subscriptions`, `customers-list-licenses`, `customers-generate-billing-links`                               |
| **Subscriptions**    | 7     | `subscriptions-get`, `subscriptions-search`, `subscriptions-update`, `subscriptions-upgrade`, `subscriptions-cancel`, `subscriptions-pause`, `subscriptions-resume`                                                                 |
| **Transactions**     | 3     | `transactions-get-by-id`, `transactions-search`, `transactions-refund`                                                                                                                                                              |
| **Discounts**        | 4     | `discounts-create`, `discounts-get`, `discounts-search`, `discounts-delete`                                                                                                                                                         |
| **Licenses**         | 4     | `licenses-activate`, `licenses-deactivate`, `licenses-validate`, `licenses-list-instances`                                                                                                                                          |
| **Customer credits** | 11    | `customer-credits-create-account`, `-get-account`, `-get-account-balance`, `-list-accounts`, `-list-entries`, `-credit-account`, `-debit-account`, `-freeze-account`, `-unfreeze-account`, `-reverse-transaction`, `-close-account` |
| **Affiliates**       | 3     | `affiliates-list`, `affiliates-retrieve`, `affiliates-list-commissions`                                                                                                                                                             |
| **Stats**            | 1     | `stats-get-summary`                                                                                                                                                                                                                 |
| **Moderation**       | 1     | `moderation-screen-prompt`                                                                                                                                                                                                          |

## Security

<Warning>
  An API key grants an agent **every** tool, including destructive ones — `transactions-refund`, `subscriptions-cancel`, `products-archive`, `discounts-delete` and `customer-credits-debit-account`. There are no read-only or scoped keys today.
</Warning>

Two things to do about that:

**Develop against test mode.** Use a test key while you build. Nothing you break is real.

**Mount only the tools you need.** `--tool` is repeatable and limits what the server exposes:

```bash theme={null}
npx -y --package creem -- mcp start \
  --api-key creem_test_your_key --server test \
  --tool products-search --tool customers-list --tool stats-get-summary
```

This is a client-side allowlist, not a credential restriction — anyone who can edit the config can remove it. Treat the API key itself as the real boundary.

## Options

| Flag           | Description                                     | Default |
| -------------- | ----------------------------------------------- | ------- |
| `--api-key`    | Your Creem API key                              | —       |
| `--server`     | `prod` or `test` — **use `test` for test keys** | `prod`  |
| `--server-url` | Override the API base URL entirely              | —       |
| `--transport`  | `stdio` or `sse`                                | `stdio` |
| `--port`       | Port, when `--transport sse`                    | `2718`  |
| `--tool`       | Mount only this tool; repeatable                | all 49  |
| `--log-level`  | `debug`, `info`, `warning`, `error`             | `info`  |
| `--env`        | Extra environment variables for the server      | —       |

## Verifying it works

<Warning>
  **A green "Connected" does not mean it's configured correctly.** `claude mcp list` — and the equivalent status in other clients — only confirms the server process started and completed the MCP handshake. It runs before any API call, so it passes even when the environment is wrong.

  A test key without `--server test` reports `✔ Connected` and then fails `401` on the first tool call.
</Warning>

Confirm with a real call instead. Ask your agent something harmless and read-only:

```
List my Creem products
```

If products come back, the key, environment and tools are all wired up. If you get a 401, see below.

## Troubleshooting

<AccordionGroup>
  <Accordion title="Every call returns 401 Invalid API Key">
    You're using a test key without `--server test`, so requests go to production. Add the flag and restart your MCP client.

    Note that your client will still have shown the server as connected — the health check can't see this.
  </Accordion>

  <Accordion title="The server doesn't start">
    Check `node --version` — v22 or newer is required. Then run the command directly in a terminal; MCP clients often swallow startup errors.
  </Accordion>

  <Accordion title="My agent can't see the tools">
    Restart the MCP client after editing its config — most read it only at launch. Then confirm the server is listed and connected in your client's MCP settings.
  </Accordion>

  <Accordion title="I want to see what the agent is doing">
    Start with `--log-level debug` and check your client's MCP logs.
  </Accordion>
</AccordionGroup>

## Related

* [Skill files](/ai/for-agents/skill-files) — reference material your agent reads
* [CLI for agents](/ai/for-agents/cli) — scriptable, `--json` output
* [TypeScript SDK](/code/sdks/typescript) — the package this server ships in