MCP server for AI agents

The MCP server at https://mcp.quellenkontor.dev/v1 is a remote server implementing the Model Context Protocol. It speaks Streamable HTTP: every request is a JSON-RPC call over POST, and the response comes back as JSON. It keeps no sessions; every request stands on its own.

• There is one tool per dataset, named after the pattern hr_datensatz.
• All tools are read-only. They are marked as read-only, non-destructive and idempotent, so many clients call them without asking for confirmation.
• Every result starts with a ready-made sentence for citing, followed by the data as text and as a structured object with the same fields as the REST API.
• The server gives the agent a set of rules: look up values instead of guessing, ask for missing inputs, cite the legal basis and source. The same text is available as the prompt quellenkontor_regeln, and below for copying.

You need a key from your account. It is sent as the header Authorization: Bearer qk_live_….

Claude Code

With one command in the terminal:

$ claude mcp add --transport http quellenkontor \
    https://mcp.quellenkontor.dev/v1 \
    --header "Authorization: Bearer $QK_KEY"

Or as a file .mcp.json in the project. Claude Code substitutes ${QK_KEY} from your environment, so the key never appears in the file:

{
  "mcpServers": {
    "quellenkontor": {
      "type": "http",
      "url": "https://mcp.quellenkontor.dev/v1",
      "headers": {
        "Authorization": "Bearer ${QK_KEY}"
      }
    }
  }
}

Cursor

In ~/.cursor/mcp.json. This file lives in your home directory, not in the project. Enter your key directly, and never put this file in a repository:

{
  "mcpServers": {
    "quellenkontor": {
      "url": "https://mcp.quellenkontor.dev/v1",
      "headers": {
        "Authorization": "Bearer qk_live_DEIN_SCHLUESSEL"
      }
    }
  }
}

VS Code

In .vscode/mcp.json. VS Code asks for the key on first launch and stores it securely:

{
  "inputs": [
    {
      "type": "promptString",
      "id": "quellenkontor-key",
      "description": "Quellenkontor-API-Schlüssel",
      "password": true
    }
  ],
  "servers": {
    "quellenkontor": {
      "type": "http",
      "url": "https://mcp.quellenkontor.dev/v1",
      "headers": {
        "Authorization": "Bearer ${input:quellenkontor-key}"
      }
    }
  }
}

Claude Desktop

Claude Desktop connects to remote servers with a fixed header through the mcp-remote bridge, which starts via npx (Node.js required). In Claude Desktop's configuration file:

{
  "mcpServers": {
    "quellenkontor": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "https://mcp.quellenkontor.dev/v1",
        "--header",
        "Authorization:${AUTH_HEADER}"
      ],
      "env": {
        "AUTH_HEADER": "Bearer qk_live_DEIN_SCHLUESSEL"
      }
    }
  }
}

The header has no space after the colon; mcp-remote reads the key from the AUTH_HEADER variable. This file lives on your machine; never put it in a repository.

Other clients

We have tested Claude Code, Cursor and VS Code. Any other client that can connect to remote servers over Streamable HTTP with a custom header will work with just these two pieces of information: the address https://mcp.quellenkontor.dev/v1 and an Authorization header carrying your key.

Claude in the browser, Claude Desktop, the Claude apps and ChatGPT add remote MCP servers as connectors. Instead of a key in the header, they use OAuth: you allow access once in the browser, and the application then queries with its own key for your account. You don't need a key from your account for this.

Claude

1. In Claude, open the connectors and add your own connector (custom connector).
2. Enter https://mcp.quellenkontor.dev/v1 as the address. Leave the client ID and client secret empty; Claude registers itself.
3. If the dialog asks how people sign in, choose sign-in with OAuth, not access without sign-in.
4. Click Connect. quellenkontor.dev opens: sign in, check the application, redirect and account, and click Allow.

The connector then works in Claude in the browser, in Claude Desktop and in the apps. In Team and Enterprise organizations, someone with admin rights adds it, and each member then connects their own account.

ChatGPT

1. In ChatGPT, turn on developer mode in the settings under Apps and create your own app with an MCP server.
2. Enter https://mcp.quellenkontor.dev/v1 as the address and choose OAuth as the authentication. ChatGPT registers itself; you don't need a client ID or secret.
3. ChatGPT then opens quellenkontor.dev. Sign in and click Allow.

What happens when you sign in

• The consent page shows the application's name, the address Quellenkontor sends you back to, your account and your plan. The application provides its own name, so rely on the address.
• When you click Allow, Quellenkontor creates an API key named “MCP: application name” and hands it to the application. It appears in your account next to your other keys and counts toward your plan's maximum number of keys.
• Tool calls through the connector use the same quota as any other request. initialize and tools/list also work without signing in; only a tool call needs the key.
• The key doesn't expire. If you revoke it in your account, the server answers the next tool call with 401. Claude then asks you to sign in again; in ChatGPT, you reconnect the app.
• Every connection creates its own key. When you reconnect an application, revoke the old key in your account, otherwise you'll quickly reach the maximum.

Technical details

The full input schema for each tool is in the dataset's reference page under "Schema for MCP". An agent gets it via tools/list.

A bot built for one topic rarely needs all 24 tools. With an addition to the address, tools/list returns only part of them, which saves tokens and prevents mix-ups:

The areas are called lohn (Wages and thresholds), sozialversicherung (Social security and levies), steuer (Tax and travel expenses), fristen (Employment law and deadlines), kalender (Calendar).

Many chatbot frameworks do not pass the server's rules on to the model. In that case, copy them into your application's system prompt. The server's rules are written in German, since that is what it actually sends:

Quellenkontor liefert amtliche HR-Werte und gesetzliche Berechnungen nur für Deutschland, Tabellen ab 2015 (Mindestausbildungsvergütung ab Ausbildungsbeginn 2020), Feiertage und Arbeitstage von 2015 bis 2035. Für Österreich, die Schweiz oder andere Länder gelten sie nicht. Rufe ein Werkzeug auf, sobald eine Antwort einen dieser Werte braucht, auch wenn du ihn zu kennen glaubst. Rechne relative Angaben wie „nächstes Jahr“ oder „ab Juli“ in ein Datum im Format JJJJ-MM-TT um und nenne es in der Antwort. Fehlen Angaben, von denen das Ergebnis abhängt, etwa Anschaffungsdatum, Geburtsdatum oder Jahrgang, Ausbildungsbeginn oder Bundesland, frag nach, statt sie zu raten. Fragt jemand nach einem ganzen Jahr, prüfe unterjährige Änderungen mit hr_verlauf und den Parametern von und bis. Übernimm das zitat, wenn es die Frage beantwortet, und nenne immer die Quelle mit URL. Geht es um einen einzelnen Bestandteil, etwa Aufmerksamkeiten oder den Pflegemindestlohn für Fachkräfte, frag mit dem Parameter bestandteil ab oder nimm Wert, gueltig_ab, rechtsgrundlage und quelle dieses Eintrags aus bestandteile. Steht vorlaeufig: true in der Antwort, sag, dass der Wert für diesen Stichtag noch nicht amtlich festgelegt ist. Meldet ein Werkzeug kein_wert, ist der Wert noch nicht verkündet: Sag das, statt zu schätzen. Gib warnungen und einschränkende hinweise weiter. Steuerfreie Höchstsätze begründen keinen Anspruch, und ein geldwerter Vorteil ist nicht die Steuer. Die Rechner wenden die gesetzliche Regel schematisch an und ersetzen keine Rechtsberatung.

This is what a tool call looks like on the wire. Normally your client handles this; for testing, curl also works:

curl -X POST https://mcp.quellenkontor.dev/v1 \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -H "Authorization: Bearer $QK_KEY" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"hr_mindestlohn","arguments":{"datum":"2027-01-15"}}}'

{
  "jsonrpc": "2.0",
  "id": 1,
  "result": {
    "content": [
      {
        "type": "text",
        "text": "Gesetzlicher Mindestlohn in Deutschland am 15.01.2027: 14,60 Euro brutto je Stunde, gültig ab 01.01.2027. Minijob-Grenze: 633 Euro im Monat. Rechtsgrundlage: Fünfte Mindestlohnanpassungsverordnung vom 5. November 2025, Stufe 2. Quelle: Fünfte Mindestlohnanpassungsverordnung vom 5. November 2025 (BGBl. 2025 I Nr. 268), https://www.recht.bund.de/bgbl/1/2025/268/VO.html. Daten: Quellenkontor (quellenkontor.dev)."
      },
      {
        "type": "text",
        "text": "(the same data as structuredContent, as compact JSON)"
      }
    ],
    "structuredContent": {
      "datensatz": "mindestlohn",
      "datum": "2027-01-15",
      "mindestlohn_brutto_stunde": 14.6,
      "minijob_grenze_monat": 633,
      "gueltig_ab": "2027-01-01",
      "gueltig_bis": null,
      "rechtsgrundlage": "Fünfte Mindestlohnanpassungsverordnung vom 5. November 2025 (BGBl. 2025 I Nr. 268), Stufe 2",
      "quelle": {
        "titel": "Fünfte Mindestlohnanpassungsverordnung vom 5. November 2025 (BGBl. 2025 I Nr. 268)",
        "url": "https://www.recht.bund.de/bgbl/1/2025/268/VO.html"
      },
      "minijob_rechtsgrundlage": "§ 8 Abs. 1a SGB IV (Mindestlohn × 130 ÷ 3, aufgerundet auf volle Euro)",
      "minijob_quelle": {
        "titel": "Bekanntmachung der Geringfügigkeitsgrenze nach § 8 Absatz 1a SGB IV vom 7. November 2025 (BAnz AT 20.11.2025 B1)",
        "url": "https://www.bundesanzeiger.de/pub/publication/VaPRG6wPGXte9CM3jI4/content/VaPRG6wPGXte9CM3jI4/BAnz%20AT%2020.11.2025%20B1.pdf"
      },
      "vorheriger_wert": {
        "gueltig_ab": "2026-01-01",
        "mindestlohn_brutto_stunde": 13.9,
        "minijob_grenze_monat": 603
      },
      "naechster_wert": null,
      "hinweise": [
        "Die Minijob-Grenze gilt je Monat. Gelegentliches unvorhersehbares Überschreiten ist in engen Grenzen erlaubt (§ 8 Abs. 1b SGB IV).",
        "Die Pauschalabgaben des Arbeitgebers für Minijobs (Kranken- und Rentenversicherung, Pauschsteuer, Umlagen) liefert der Datensatz minijob-abgaben (/v1/hr/minijob-abgaben)."
      ],
      "lizenz": "Werte aus amtlichen Werken (§ 5 UrhG). Nutzung der Zusammenstellung nach den Nutzungsbedingungen von Quellenkontor.",
      "stand": "2026-09-23",
      "zitat": "Gesetzlicher Mindestlohn in Deutschland am 15.01.2027: 14,60 Euro brutto je Stunde, gültig ab 01.01.2027. Minijob-Grenze: 633 Euro im Monat. Rechtsgrundlage: Fünfte Mindestlohnanpassungsverordnung vom 5. November 2025, Stufe 2. Quelle: Fünfte Mindestlohnanpassungsverordnung vom 5. November 2025 (BGBl. 2025 I Nr. 268), https://www.recht.bund.de/bgbl/1/2025/268/VO.html. Daten: Quellenkontor (quellenkontor.dev).",
      "datenstand": "2026-09-23.1"
    },
    "isError": false,
    "_meta": {
      "quellenkontor.dev/kontingent": {
        "limit": 1000,
        "verbraucht": 12,
        "tarif": "kostenlos"
      }
    }
  }
}

A tool call that returns a result counts like one REST API request. initialize, tools/list and calls with invalid arguments are free. You can find the state after every call in the result under _meta, in the key quellenkontor.dev/kontingent, with limit, usage and plan.

• Supported protocol versions: 2025-11-25, 2025-06-18, 2025-03-26, 2024-11-05. The server responds with the version the client requests, otherwise with the latest.
• Methods: initialize, ping, tools/list, tools/call, prompts/list and prompts/get (prompt quellenkontor_regeln). resources/list returns an empty list.
• No server-sent stream: a GET on the address gets a 405 from the server; notifications without an id get 202.
• No sessions: the Mcp-Session-Id header is not needed.

The MCP Inspector is the protocol's official testing tool. It shows every tool with its schema and lets you run calls by hand:

npx @modelcontextprotocol/inspector

In the Inspector, choose the Streamable HTTP transport, enter the address https://mcp.quellenkontor.dev/v1, and under Authentication set the Authorization header to Bearer qk_live_….

A good agent states the legal basis and source from the result, asks about missing inputs and says when a value is still provisional. If your application gives the model its own instructions, add that HR values may only come from the tools.