Developer builds AI-powered news aggregator with MCP, DeepSeek, and Telegram bot

The reason for this project was not an abstract idea, but a concrete problem: on February 28, 2026, the author found himself in Dubai during the interception of Iranian drones and could not quickly get a complete picture of what was happening. This gave birth to a multilingual news aggregator that collects news, translates it, tags it, and delivers it not just to people, but to AI assistants as well.

How the Project Appeared

The author describes the starting point very pragmatically: Google and Yandex returned materials that were two hours old, while Telegram channels provided conflicting versions of events. When news is needed right now, such a delay turns from a minor inconvenience into a real problem.

At first, the solution was simple — a Telegram channel about the Middle East where RSS feeds from major world news agencies were manually collected. But manual mode quickly hit a volume ceiling. The stream grew to 200+ publications a day, and maintaining it without automation became pointless. Then came a parser, then auto-translation, then IT sources and economic news were added. In the end, a local idea for an operational channel turned into a full-fledged product: a news aggregator website with a feed, an AI chat, an API, an MCP server, and a separate Telegram bot.

What's Inside the System

The project is built as a set of independent components, not as a monolith. A separate FeedManager service pulls RSS, processes feeds, and writes data to MySQL. The Flask part reads already-processed articles and does not depend on whether the parser got stuck on some unstable source. Automatic tag annotation works as a standalone process: it reads fresh materials from the database and writes results back, while the Telegram bot responds to users in its own loop.

• 80+ sources in a single stream, including international news agencies, IT media, and business publications
• 9 thematic categories and about 50 tags for navigating the feed
• automatic translation of materials into Russian immediately after publication
• REST API with search, filtering by tags, dates, and sources
• MCP server that allows AI assistants to work with the news database as a tool

On top of the aggregator sits an API and MCP server for external clients and AI agents. A key detail is not just raw JSON output, but structured responses through Pydantic models like ArticleSummary and SearchResult. This reduces errors in processing and saves tokens. According to the author's plan, an assistant can find the right tag on its own, request articles from recent days, read full texts, and compile a digest without manual data preparation. Tools are picked up dynamically every five minutes, so new features can be added without rewriting agent logic.

Why Not Too Much AI

AI in this project is not used everywhere, only where it truly pays off. For the chat interface, DeepSeek was chosen: according to the author, it is 15–30 times cheaper than alternatives while delivering comparable quality for news search and summarization. For a free product, this is critical. The same pragmatism is evident in how translations are stored: multilingual versions of an article sit directly in a JSON field of the record, which speeds up access and simplifies scaling to new languages.

The most telling compromise is automatic tag annotation. Instead of an LLM, the author used keyword matching across 47 topics and about 600 keywords in English and Russian. This approach yields about 85% accuracy, runs in less than a millisecond per article, and consumes zero tokens. Against the backdrop of universal enthusiasm for "plug a neural network into every step," this looks like a mature engineering position: AI is needed where it adds value, not just because it looks trendy.

What This Means

This case clearly shows that a news AI product today is not just a model that can paraphrase texts. Real value appears where there is reliable data collection, proper structure, a clear API, and a protocol like MCP for connecting assistants. For media, research teams, and internal analytics services, this is a signal: from an ordinary feed you can build a working infrastructure for live digests, search, and automated answers.