Manual

🇩🇪 Diesen Artikel auf Deutsch lesen

Setting up the MCP server (connecting AI agents to easycompliance)

With the easycompliance MCP server, AI assistants such as Claude screen business partners against sanctions and PEP lists and maintain your daily monitoring list – via the open MCP standard (Model Context Protocol). This article walks you through the setup in a few minutes.

What the MCP server can do – and what it cannot

The MCP server provides your AI assistant with six tools: screening a name against the sanctions lists, screening a name against the PEP list (politically exposed persons), retrieving the hits of the last 24 hours, querying the date of the last list update, and adding names to or removing them from the daily monitoring list. Screening results are always structured (hit yes/no, hit count, full match data with a percentage match score).

What it deliberately does not do: it does not replace the professional assessment of a hit – that belongs in the customer portal at https://kunde.easycompliance.de (Monitoring → Hits). The tool descriptions also explicitly oblige the AI model never to soften or omit hits.

Step by step

1. Have your API access and API key ready

The MCP server uses the easycompliance REST API and therefore requires API access, which is booked as an add-on to your plan. You receive the API key from customer service during setup: Start a chat with customer service

2. Add the MCP server to your AI client

For local use your computer needs Node.js version 20 or later; the mcp-easycompliance package is downloaded from npm automatically on first start. In Claude Desktop, add the server to the configuration file claude_desktop_config.json (Settings → Developer → Edit configuration):

{
    "mcpServers": {
        "easycompliance": {
            "command": "npx",
            "args": ["-y", "mcp-easycompliance"],
            "env": {
                "EC_API_KEY": "YOUR-API-KEY"
            }
        }
    }
}

In Claude Code, a single terminal command is enough:

claude mcp add easycompliance --env EC_API_KEY=YOUR-API-KEY -- npx -y mcp-easycompliance

Cursor and other MCP clients use the same JSON structure as Claude Desktop or start the server via the command npx -y mcp-easycompliance with the environment variable EC_API_KEY set.

ChatGPT – unlike Claude Desktop – does not run MCP servers locally but connects them as remote connectors. In ChatGPT, open “Settings → Apps & Connectors” (in Business and Enterprise accounts the administrator enables connectors for the team) and create a new connector with these values:

  • Name: free to choose, e.g. easycompliance
  • Type: Streamable HTTP
  • URL: https://www.easycompliance.de/mcp.api
  • Bearer token: leave empty (not required)
  • Headers: add a header named X-API-Key with your API key as its value

After saving, the six tools are available directly in ChatGPT. Self-built agents using OpenAI models (OpenAI Agents SDK, Codex CLI), on the other hand, start the package locally via npx -y mcp-easycompliance, exactly as described above.

3. Test the connection

Ask your AI assistant a test question, for example: “When were the sanctions lists last updated?” If it answers with a date, the connection works. A subsequent trial screening of a name returns either “no hit” or the structured match data.

Good to know

  • Every executed tool call is a regular API request: completed screenings, list-status and hit queries as well as executed monitoring changes appear in the API log of the customer portal and count towards the API quota of your plan. Rejected requests (e.g. an invalid API key) and deletion attempts without a matching entry are not logged. You can review the usage under “API statistics” in the customer portal.
  • The API key is passed as an environment variable only and is never written to files or logged by the MCP server. Note, however, that your AI client's configuration file (e.g. claude_desktop_config.json) stores the entered key locally – protect that file like a password.
  • The two monitoring tools are marked as write tools – many MCP clients therefore ask for confirmation before executing them.
  • If your account uses customer-specific API addresses, set them via the optional environment variables EC_SANCTIONS_URL and EC_PEP_URL.
  • Without a local installation: easycompliance also provides the MCP server as a remote endpoint at https://www.easycompliance.de/mcp.api; the API key is transmitted per request in the HTTP header X-API-Key. ChatGPT uses exactly this endpoint (see step 2); it also works with any other client that can send custom HTTP headers. Details are explained on the MCP page on the website (German).

Frequently asked questions

Does the MCP server cost extra?

There are no separate fees for the MCP server itself; the package is open source (MIT license). However, it requires API access, which is booked as an add-on to your plan, and every screening counts towards the API quota – exactly like a direct API call.

Which lists does the MCP server screen?

The same ones as your easycompliance account: the sanctions lists (EU/US/HADDEX) via the check_sanctions_list tool and – if your plan includes PEP screening – the PEP list via check_pep_list.

Can the AI assistant accidentally delete entries from my monitoring list?

Deleting is only possible via the tool remove_from_monitoring, which is explicitly marked as a write tool and requires a specific name or reference; the server rejects empty delete requests. Many MCP clients additionally ask for your confirmation before running write tools.

What happens when there is a hit?

The AI assistant receives the structured result with all match data and is obliged by the tool description to report every hit in full. You then carry out the professional assessment in the customer portal at https://kunde.easycompliance.de (Monitoring → Hits).

Requirements

  • easycompliance account with booked API access (API key from customer service)
  • MCP-capable AI client, e.g. Claude Desktop, Claude Code, Cursor or ChatGPT (ChatGPT: as a remote connector, see step 2)
  • For local use: Node.js version 20 or later

Last reviewed: