German Newsfeed MCP Server

Disclaimer: This is a private, unofficial project. It is not an ARD product and is neither operated nor endorsed by ARD, ARD-aktuell, or NDR. It is developed independently of the author's professional employment. "ARD" and "tagesschau" are trademarks of their respective owners and are named here solely to describe the API being accessed.

This project merely connects the public API to an MCP-capable AI assistant. ARD-aktuell is responsible for the API itself, its operation, and its content; this project cannot provide information on any of those. Please raise any concerns about this project, in particular from rights holders, as a GitHub issue. Substantiated concerns will be addressed promptly.

In thirty seconds

This Model Context Protocol (MCP) server connects your AI assistant (Claude Desktop and others) to the public news API of tagesschau.de: current headlines, category and regional news, and full-text search, locally via stdio, no API key required.

Example: asked "Was sind die aktuellen Schlagzeilen?", the assistant answers with the current top stories from tagesschau.de, each with title, date, summary, and a link to the article.

Language:

• 🇩🇪 Deutsch
• 🇬🇧 English

What is this?

This MCP server connects the public news API of tagesschau.de to your AI assistant (Claude, Open Claw, etc.).

Once connected, your AI can answer questions like:

• "Was sind die aktuellen Schlagzeilen?"
• "Zeig mir die neuesten Wirtschaftsnachrichten."
• "Suche nach Artikeln über Ukraine."
• "Welche Regionalnachrichten gibt es aus Bayern?"

For details on the API and its terms, see Data Source and Terms of Use. No API key required.

Data Source and Terms of Use

This server calls the publicly accessible endpoint www.tagesschau.de/api2u/, operated by ARD-aktuell. The delivered content originates from the ARD broadcasters, remains subject to their rights and to the terms of use of tagesschau.de: https://www.tagesschau.de/nutzungsbedingungen/

The API is not officially documented. Community documentation is available at bund.dev (https://tagesschau.api.bund.dev). bund.dev is a civil-society documentation project, neither the operator of the API nor a rights holder of the content. Its documentation served as a reference for this project but does not grant any usage rights.

Applicable limits: at most 60 requests per hour, and no republication of the content except for offerings under a CC licence (https://tagesschau.de/creativecommons). Compliance is the responsibility of whoever operates a given instance.

The robots.txt of tagesschau.de additionally declares an express reservation of rights under Section 44b(3) of the German Copyright Act (as of 2026-05-19): text and data mining and the automated use of the content for training or fine-tuning AI models are prohibited without written consent. Expressly exempt is automated access for the sole purpose of retrieval-augmented generation (RAG) or grounding, provided the technical directives of the robots.txt are complied with and the content remains attributed to its original source. This server falls under that exemption: it passes content to the assistant exclusively together with source links. The retrieved content must not be used to train AI models.

Features

Project Structure

german-newsfeed-mcp/
├── src/
│   └── german_newsfeed_mcp/
│       ├── __init__.py      # Package metadata
│       ├── config.py        # Environment-driven configuration
│       ├── client.py        # Async HTTP client (httpx) + error handling
│       ├── rate_limiter.py  # Token-bucket rate limiter
│       ├── validators.py    # Domain constants + validation helpers
│       ├── formatters.py    # Markdown rendering of news items & channels
│       ├── tools.py         # MCP tool business logic
│       ├── resources.py     # MCP resource business logic
│       └── server.py        # Composition root: FastMCP + run() entry-point
├── tests/
│   ├── conftest.py          # Shared fixtures & mock payloads
│   ├── test_client.py       # Client unit + live integration tests
│   ├── test_rate_limiter.py # Rate-limiter unit tests
│   ├── test_validators.py   # Validator unit tests (URL & ressort checks)
│   ├── test_formatters.py   # Formatter unit tests (pure functions)
│   ├── test_tools.py        # Tool unit + live integration tests
│   └── test_resources.py    # Resource unit + live integration tests
├── main.py                  # Thin entry-point (calls server.run())
├── pyproject.toml           # Project metadata, deps, pytest & pylint config
├── uv.lock                  # Locked dependency graph (commit this!)
├── Dockerfile               # Multi-stage production image
├── docker-compose.yml       # One-command remote deployment
├── .env.example             # Configuration template
├── Makefile                 # Developer shortcuts (test, lint, docker, clean)
├── CHANGELOG.md             # Version history
├── CONTRIBUTING.md          # How to contribute
└── README_de.md             # German version of this file

Quick Start Local (Claude Desktop)

Also applicable to other AI assistants, edit their respective config instead. This mode uses stdio transport; the server is launched as a child process. No port is needed.

Prerequisites

• macOS / Linux / Windows (WSL2)
• Python 3.12+
• uv: curl -LsSf https://astral.sh/uv/install.sh | sh
• Claude Desktop

Step 1: Clone and install

git clone https://github.com/Jakolo121/german-newsfeed-mcp.git
cd german-newsfeed-mcp
uv sync

Step 2: Verify it works

uv run python -c "from german_newsfeed_mcp.server import mcp; print('OK!', mcp.name)"
# Expected: OK! German Newsfeed MCP

Step 3: Connect Claude Desktop

Open your Claude Desktop config file:

Add the following entry (adjust the path to your clone):

    "german-newsfeed": {
      "command": "uv",
      "args": [
        "--directory",
        "/absolute/path/to/german-newsfeed-mcp",
        "run",
        "german-newsfeed-mcp"
      ]
    },

Step 4: Restart Claude Desktop

Quit and reopen the application, or reload its MCP servers, depending on the application.

Step 5: Try it!

Ask your assistant:

"Was sind die aktuellen Nachrichten?"

Remote / Docker Deployment

Self-hosting for your own or team-internal use. The recommended default is stdio (see Quick Start); the streamable-http transport is an option you choose deliberately.

Security note: The HTTP transport has no authentication. Do not expose it publicly without an authentication layer in front (e.g. a reverse proxy). The Compose setup therefore deliberately binds the port to 127.0.0.1 only. Whoever makes an instance reachable for third parties becomes the responsible operator under the terms of use.

Rate limiter limitations: stateless_http=True refers to MCP sessions, not to the rate limiter. The token bucket is process-local in-memory state. Two limitations follow:

1. Multiple replicas against the same upstream API multiply the request budget.
2. A container restart resets the bucket to full. Combined with restart: unless-stopped and a crash loop, this can exceed the limit. Watch the logs.

(The legacy sse transport is still supported.)

Prerequisites

• Docker 24+
• Docker Compose v2+

Step 1: Create your .env file

cp .env.example .env
# Edit .env if you want a different port or log level

Step 2: Build and start

docker compose up --build -d

Or:

make docker-build
make docker-run

The server starts at http://localhost:8000.

Step 3: Verify health

docker compose logs german-newsfeed-mcp
docker compose ps

Or:

make docker-logs

Step 4: Connect Claude Desktop (Streamable HTTP)

    "german-newsfeed": {
      "command": "npx",
      "args": ["mcp-remote", "http://localhost:8000/mcp"]
    },

For external servers: put an authentication layer in front first (see the security note above), then adjust the loopback binding in docker-compose.yml and replace localhost with your proxy's address.

Stop / Update

docker compose down
docker compose up --build -d

Configuration Reference

All settings are read from environment variables (or a .env file).

The defaults apply when running directly (uv run german-newsfeed-mcp). The Compose setup overrides them: it sets HOST=0.0.0.0 and PORT=8000 (see docker-compose.yml).

Available Tools

These tools are callable by your AI assistant.

get_latest_news

Get the top stories.

Returns the full article text for each item (homepage feed). Every item also carries a 🔗 Volltext: link that get_article accepts.

get_article

Get the complete text of a single article.

Takes the 🔗 Volltext: link printed with every item by get_latest_news, get_news_by_ressort and get_regional_news; a plain https://www.tagesschau.de/<path>.html article link works too.

Costs exactly one upstream request per call out of the 60/h budget — call it for the articles that matter, not for every headline in a listing.

search_news results carry no article link and cannot be passed to get_article. Locate the article via get_latest_news, get_news_by_ressort or get_regional_news first.

get_news_by_ressort

Filter news by category.

Returns metadata only — title, topline, date and a teaser sentence — no full article text. Every item carries a 🔗 Volltext: link that get_article resolves into the full text.

Ressort strings are automatically normalised to lowercase: "Inland", "INLAND" and "inland" are all equivalent.

get_regional_news

News from a specific German state.

Returns metadata only — title, topline, date and a teaser sentence — no full article text. Every item carries a 🔗 Volltext: link that get_article resolves into the full text; because regional items originate at an ARD state broadcaster, they additionally carry a 📰 Quelle: line naming that broadcaster.

search_news

Full-text search across all available articles.

Returns metadata only — title, date and article type — no full article text. Search results carry no article link and therefore cannot be passed to get_article.

get_channels

List all live channels with stream URLs.

(No parameters)

Available Resources

Resources are addressable URIs that MCP clients can read directly.

Development Guide

Setup

git clone https://github.com/Jakolo121/german-newsfeed-mcp.git
cd german-newsfeed-mcp
uv sync --extra dev

Start the server:

uv run german-newsfeed-mcp

Or:

make run

Code organisation (SOLID)

Each module has exactly one responsibility:

Adding a new tool

1. Add a tool_<name>() async function in tools.py
2. Register it with @mcp.tool() in server.py
3. Write unit + integration tests in tests/test_tools.py

Running the Tests

Unit tests (no internet required, fast)

uv run pytest -m "not integration"     # run all unit tests
uv run pytest -m "not integration" -v  # verbose output
uv run pytest tests/test_formatters.py # single file

Or:

make test

Live integration tests (requires internet)

uv run pytest -m integration           # all live tests
uv run pytest -m integration -v        # verbose

Full suite

uv run pytest

Or:

make test-all

Quality gate (lint + tests)

uv run pylint src/german_newsfeed_mcp/
uv run pytest

Or:

make check

Expected results

251 passed            ← unit tests (no network)
 22 selected          ← integration tests (live API)

Makefile Reference

make test          # fast unit tests (no network, ~0.3 s)
make test-all      # unit + live integration tests
make lint          # pylint
make check         # lint + unit tests — use as CI gate
make run           # start server in stdio mode (Claude Desktop)
make run-http      # start server in streamable-http mode
make docker-build  # build Docker image
make docker-run    # docker compose up -d
make docker-stop   # docker compose down
make docker-logs   # tail docker compose logs
make clean         # remove __pycache__, .pytest_cache, dist, etc.

Troubleshooting

Claude Desktop shows no MCP tools

1. Check that the claude_desktop_config.json path is absolute
2. Run uv run python main.py in the terminal, it should start without errors
3. Fully quit and reopen Claude Desktop (Cmd+Q, not just close window)

Docker container exits immediately

docker compose logs german-newsfeed-mcp

Or:

make docker-logs

Common causes: wrong TRANSPORT value (must be streamable-http in Docker), port already in use.

API timeouts

The upstream API occasionally rate-limits certain endpoints. This is normal, the server returns a descriptive error message rather than crashing. Retry after a few seconds.

Rate-limit errors

If the server reports "Rate limit exceeded", the local request budget (RATE_LIMIT_PER_HOUR, default 60/h) is exhausted. No request was sent upstream. Try again later.

Import errors in tests

uv sync --extra dev     # ensure dev deps are installed
uv run pytest           # always run via uv, not bare pytest

When it stops working

The upstream API is not officially documented and can change without notice. You can tell by the tools suddenly returning empty lists or error messages although tagesschau.de is reachable, and by the live integration tests failing (uv run pytest tests/ -m integration).

All endpoints are defined in a single place: ENDPOINTS in src/german_newsfeed_mcp/client.py. API changes can be tracked there.

The CI job upstream-check (.github/workflows/ci.yml) runs exactly these live tests weekly against the real API and fails loudly when the response format no longer matches.

License & Acknowledgements

Apache License 2.0.

The delivered news items are content of the ARD broadcasters, subject to their rights and the terms of use of tagesschau.de, see Data Source and Terms of Use.

Thanks to AndreasFischer1985, the bund.dev community for documenting the API and above all to the journalists at the ARD broadcasters, whose work this project merely passes along.