mcp-news

News for the reader — not news that turns the reader into the product.

Your news feed is currently chosen by someone whose income depends on how long you keep scrolling. mcp-news inverts that. You write down what you are interested in, and that is the ranking algorithm — a list of rules you can read, stored in a file you own. Nothing about you is transmitted anywhere, because there is nowhere to transmit it to: the whole thing runs on your own machine.

You set it up in a browser. There is no file to edit, no environment variable to export and no terminal to open on the way to a working feed. The configuration is plain YAML because it is yours, not because you are expected to type it.

Ask it questions through any MCP-compatible AI assistant, or open the dashboard on your phone. Both read the same corpus you collected, ranked the way you said.

What it is

A self-hosted news aggregator with three faces:

Related MCP server: rss-news

What makes it different

• Your algorithm is a list of rules you wrote. You edit it in the browser; it is stored as config/profile.yaml, which you can read, copy and keep under version control. Change it and the feed reorders immediately. Read it and you know exactly why an article surfaced — the dashboard shows the matched rules as chips on every item.

• Runs without any AI at all. No API key, no GPU, no model? It still collects, deduplicates, stores and keyword-searches everything, and ranks it against your profile. Add a model later and it enriches what it already collected. Nothing is lost by starting simple.

• Four provider slots, two chains, automatic failover. Two local/private and two cloud engines by default, for chat and embeddings. If one is down, the next in the chain takes over. Add more; the registry doesn't care.

• Sources live in dated, versioned files. Feeds die, move and get replaced. config/sources/*.yaml records when each was added, last verified, and whether it has been superseded — so a stale list is visible instead of silent.

• Speaks your language, and adding one is a single file. Every string lives in a flat JSON catalogue at web/i18n/<lang>.json. English and Portuguese ship complete; a new language is a copied file, translated, with no build step and no restart. Right-to-left is a one-line change. See docs/LOCALIZATION.md.

• Sharing sends the publisher's link, never yours. Every article carries a Share control that opens your device's own share sheet, so it reaches every application you actually have installed. Browsers without one get a short list of plain share links instead. See Sharing below.

• No third-party UI framework. The dashboard is hand-written HTML, CSS and JavaScript served by our own API. No vendor banner, no telemetry, no analytics, no build step, and no Node anywhere in the container.

Quick start

git clone https://github.com/RaymondJPayne/mcp-news
cd mcp-news
docker compose up -d

Open http://localhost:8378. A setup wizard asks you four things — your language, where to keep your articles, which source bundles to read, and what you are interested in — and then starts collecting. The first articles usually appear within a minute or two.

Nothing above needs a file, an editor or a terminal beyond that one command. There is no .env to copy and no example config to rename.

Adding a model, later, if you want one

Entirely optional. With no model configured the application collects, de-duplicates, stores, keyword-searches and ranks — see Capability levels below. To add one, open Settings → AI models and point a slot at any OpenAI-compatible endpoint, local or hosted. Keys are never stored in a configuration file: the file names an environment variable, and you put the key in .env or your shell environment.

LOCAL_CHAT_BASE_URL=http://host.docker.internal:1234/v1
LOCAL_EMBED_BASE_URL=http://host.docker.internal:1235/v1
# and/or
CLOUD_CHAT_API_KEY=sk-...
CLOUD_EMBED_API_KEY=sk-...

Then docker compose exec node mcpnews enrich --backlog processes everything you already collected, highest interest score first.

Your algorithm is a list of rules

You write these in the browser. This is what they look like on disk.

interests:
  - name: Semiconductor policy
    match: [export control, fab, lithography, ASML, TSMC]
    weight: 5
  - name: AI governance
    match: [AI Act, model regulation, algorithmic accountability]
    weight: 4

places:
  - name: Brazil
    weight: 5
  - name: Ontario
    weight: 4

mute:
  domains: [example-clickbait.com]
  keywords: [horoscope]

No training, no feedback loop you cannot see, no drift. If it surfaces something odd, the dashboard tells you which rule did it.

Sharing

Every article in Today and in the article view has a Share control. What goes out is the publisher's own link — never a link to this machine, never an address only your network can resolve, and never an id from your database. An article we collected without a source link shows no Share control at all.

Where the browser has the Web Share API — every current mobile browser and most desktop ones — one tap opens the operating system's share sheet, which reaches every application you have installed. That is the whole feature, and it is better than any list we could write. Browsers without it get a short explicit list instead: Mastodon, Bluesky, LinkedIn, Reddit, WhatsApp, Telegram, Facebook, X, email, and Copy link. Each of those is a plain share-intent URL opened in a new tab — no vendor SDK, no embedded widget, no remote script, no tracking pixel — and the icons are inline SVG we drew ourselves. Nothing about a share is recorded. Your Mastodon server is remembered by your browser alone; Settings shows it and can forget it.

Attribution. A share carries one optional extra line crediting where the article reached you. It is on by default and configurable in Settings → Sharing: turn it off, change the wording, or point the link at your own page instead of the repository. Turning it off turns it off — nothing here re-enables it later. LinkedIn, Facebook and Reddit accept only a link and write their own preview, so the credit line does not travel with those three; the Settings screen says so.

Capability levels

The application announces which level it is running at, in the header and in /api/status, so a degraded mode is visible rather than silently worse.

Level 0 is the floor the project guarantees, and it is a real product rather than a placeholder.

Connecting an AI assistant

{
  "mcpServers": {
    "news": {
      "command": "docker",
      "args": ["compose", "exec", "-T", "node", "mcpnews", "mcp"]
    }
  }
}

Then ask: "What changed on semiconductor export controls in the last fortnight?"

See docs/MCP.md for the full tool surface.

Documentation

Status

Collecting, ranking, the dashboard and the MCP server work end to end with no model configured. Embeddings, chains and failover work when a provider is configured. Translation, entity extraction and change detection do not exist yet and say so where you would expect to find them.

docs/ROADMAP.md states each phase honestly, including what is stubbed and why.

This project ships no authentication and claims none. It binds to 127.0.0.1. If you want it on a network, that is your reverse proxy's job.

Licence

AGPL-3.0-or-later for the code — see LICENSE.

The source-file schema, the shipped source bundles and the interface string catalogues are MIT instead, so that writing a compatible client, adding a language or reusing the source list never means taking on AGPL obligations. Interoperability should not require adopting our licence. The exact file list is in LICENSE-MIT.