← Blog · October 3, 2026

CCO MCP Server now available: maintaining the SAP Customer Checkout Manager with AI assistants

Much of the day-to-day work in the SAP Customer Checkout Manager is routine maintenance: creating the articles for a new menu, setting prices in several price lists, building quick selections, checking day-end closings for deviations. None of these tasks is complex, but together they take a considerable amount of time in the Manager's web interface.

The CCO MCP Server allows this work to be delegated to an AI assistant such as Claude, Codex or ChatGPT. The task is described in plain language, and the assistant carries it out in the Manager. The server is now available in the customer portal. A registered portal account is all that is required, and the server is free of charge.

What MCP is

The Model Context Protocol (MCP) is an open standard for connecting AI assistants to other systems. An MCP server provides a set of tools, for example "search articles" or "set price". The assistant reads the tool descriptions and decides which tools to call for a given request. MCP is supported by a growing number of assistants and agent tools, including Claude Code, Claude Desktop, Codex and ChatGPT Desktop.

The CCO MCP Server exposes the REST API of the Manager as such tools. It uses the same API as the Manager's web interface, so every change is stored in the Manager and can be reviewed there as usual. Both editions that are current today are supported:

SAP Customer Checkout 2.0 FP21 (on-premise Manager)
SAP Customer Checkout, cloud edition 3.x

The server is not tied to a particular assistant. It runs locally as an MCP server over stdio and therefore works with any client that can start such a server. The documentation describes the setup for Claude Code, Claude Desktop, Codex and ChatGPT Desktop.

Capabilities

The server provides 48 tools. Users do not need to know the tool names, since the assistant selects the appropriate tools itself. They cover the following areas:

Articles and prices. Search, read, create with prices in several price lists and tax assignment, update, set a price per price list.
Quick selections. Create structures, add category tabs, place article, function and text buttons, upload button images, copy a complete quick selection, assign it to org nodes.
Users and organization. Create POS users with role and barcode login, create org nodes, assign table areas.
Reporting. Sales reports such as revenue per POS or per article, day-end closings including their deviations, receipts.
POS actions. Deploy an action jar to the tills in a single call and track the rollout per till.
Receipt forwarding. Set up a forwarding of receipts to an HTTP endpoint and switch it on or off.

For cases without a dedicated tool, ccom_request sends a plain request to the Manager API.

Typical requests look like this:

Create the article "Falafel wrap" at 7.50 EUR in the standard price list,
article group Snacks, 19% VAT.

Raise all prices in the article group Hot beverages by 20 cents.

Copy the quick selection of the Nord branch and assign the copy
to the Süd branch.

How much revenue did we make today per POS?

Show this week's day-end closings. Were there any deviations?

A single server instance can serve several Managers at the same time, including a mix of on-premise and cloud systems. This allows requests that span systems, for example "Read the quick selection 'Bar' from the test system and rebuild it on the cloud Manager".

Setup

Requirements:

A portal account. Registration at portal.stefanbaust.de is free of charge.
Access to a Manager: a Manager user for FP21, or an API client for the cloud edition.
An AI assistant that supports local MCP servers, with its own license or subscription, for example Claude Code, Claude Desktop, Codex or ChatGPT Desktop.
Linux x64, Windows x64 or macOS (Intel or Apple Silicon). The server is a single executable and requires neither Node.js nor Java.

Step 1: Run the installer

# Linux / macOS
curl -fsSL https://portal.stefanbaust.de/api/v1/install/cco-mcp-server.sh | sh

# Windows (PowerShell)
irm https://portal.stefanbaust.de/api/v1/install/cco-mcp-server.ps1 | iex

To review the script before running it, download it first and open it in an editor.

The installer displays a short code and a link to the portal. After you sign in and confirm the code, the installer downloads the executable for your platform, verifies its SHA-256 checksum and retrieves your license. All files are stored in ~/.cco-mcp-server/, and the cco-mcp-server command is added to your PATH. On Linux and macOS, open a new terminal afterwards.

Step 2: Add a Manager

cco-mcp-server add-manager

The wizard asks for a profile key such as fp21-test, a display name, the edition, the URL, the credentials and whether the profile should be read-only. It tests the connection before saving. The credentials are stored in ~/.cco-mcp-server/managers.json, readable only by your user account, and not in the configuration of the assistant. Run the wizard again for each additional Manager.

For the cloud edition, client ID and secret are available in the Manager under Configuration → General → API settings. On the first API call, the Manager creates a technical user named after the client ID. This user must be assigned a role once, for example Administrator. Until then, every call returns HTTP 401. The connection test of the wizard points this out.

Step 3: Register the server in the assistant

All clients need the same information. The executable is started as a local (stdio) MCP server without arguments, which makes all saved Manager profiles available.

Claude Code:

claude mcp add cco-manager -- ~/.cco-mcp-server/bin/cco-mcp-server

Codex CLI:

codex mcp add cco-manager -- ~/.cco-mcp-server/bin/cco-mcp-server

Desktop applications require the full path to the executable. For Claude Desktop, the entry goes into claude_desktop_config.json:

{
  "mcpServers": {
    "cco-manager": {
      "command": "/Users/you/.cco-mcp-server/bin/cco-mcp-server"
    }
  }
}

In the Codex and ChatGPT desktop applications, add an MCP server of type STDIO in the settings and enter the same path. Menu names change between versions, so the documentation describes the current steps. On Windows, the executable is located at C:\Users\you\.cco-mcp-server\bin\cco-mcp-server.exe.

Then restart the client, open a new conversation and ask:

Which CCO Managers are configured, and are they reachable?

Data flow

Data flow: the AI assistant and the CCO MCP Server run on your computer. The MCP server communicates directly with the CCO Manager. The AI model at your provider receives prompts and tool results. The portal is used only for installation, updates and the license.

The server runs on your computer and communicates directly with the Manager. The portal is only involved in installation, updates and licensing. Article, receipt and user data never pass through it.

The AI model is a different matter. Everything the assistant reads through the tools is sent to the model provider like any other input, which is Anthropic for Claude and OpenAI for Codex and ChatGPT. Receipts, user lists and customer records can contain personal data, and a read-only profile does not change that. Where the GDPR applies, check which data you may share with your AI provider and whether a data processing agreement (DPA, AVV in German) is required.

Version 1.1.0 introduced an experimental privacy filter, which is disabled by default. With "privacyMode": "minimize" set on a profile, it removes known personal fields and free-text notes from the responses before they reach the assistant. Revenue, article, price and quick selection workflows remain available. Some functions, such as raw API requests and action deployment, are blocked while the filter is active. The filter does not guarantee anonymity, since article names, labels and fields unknown to the filter can still contain personal data. This section is technical information, not legal advice.

Control over changes

The server acts with the permissions of the configured user or API client and can modify master data. I recommend the following practices:

Evaluate the server on a test Manager first.
Review tool calls before approving them. Claude Code and Claude Desktop request approval for every tool call by default. Check how your client handles approvals, and do not approve write tools permanently.
Use read-only profiles for reporting. The server then rejects all write tools itself, regardless of the client settings.
Create a dedicated Manager user with the minimum permissions required instead of storing an administrator account.
Use deploy_action with care. It distributes a jar that runs inside the POS on every assigned till, which amounts to a software rollout to your stores. On a production Manager, this should be a deliberate decision.

License and warranty

The CCO MCP Server is free of charge for every registered portal account, including the write tools and multiple Manager profiles. cco-mcp-server update installs new versions and renews the license at no cost. The license is verified offline.

The server is provided as is, without any warranty. There is no guaranteed support and no contractual entitlement to future releases or fixes. Users decide which actions the server may perform on their systems and are responsible for them. Changes should be tested on a test Manager before they are applied to production. The terms are set out in the EULA in the portal. The server is free to use, but it is not open source. The portal provides a compiled executable for each platform.

Further development

The CCO MCP Server is being extended continuously. Feature requests are welcome. If a function for your use case is missing, please send a short description to stefan.baust@stefanbaust.de.

ccosapSAPCustomerCheckoutmcpaiagents