AI News Radar: A Self-Hosted AI News Pipeline with Multi-Source Deduplication and Persona-Based Commentary

What Problem It Solves and Who It Is For

AI News Radar addresses the common pain of tracking AI developments across scattered sources: official blogs, changelogs, X posts, and aggregators that repeat the same story. The README describes a workflow where you open dozens of pages, manually filter duplicates, and guess which item matters. This project automates that first pass. It is for three groups. Ordinary readers get a web page with tabs and a curated/raw toggle. Content creators get source links and multi-source labels to compare wording and assess credibility. Developers and agents get a pipeline they can fork, with a Scout Skill (the README calls it 伯乐Skill) to help maintain sources and scraping logic. The project explicitly targets users who want to avoid API keys, login states, and LLM quotas by default.

How the Pipeline Works: From Sources to JSON

The architecture is a linear pipeline, as shown in a Mermaid diagram in the README. It starts with a source list, then a classification step that decides the source type: official RSS or changelog, private OPML or RSS, public GitHub feed or JSON, static pages with Jina fallback, or AgentMail email subscriptions. High-risk sources are skipped. After classification, the pipeline fetches and structures content, then deduplicates and normalizes it. Next comes AI relevance scoring and tagging, followed by story merging with multi-source evidence. Finally, source health and AI share are computed. The output is a set of static JSON files in a data directory. The frontend reads only these files, so no backend service is needed. GitHub Pages serves as the canonical data source, while a Vercel site is just another front door. This design means the core process does not consume model quotas once it is live, as the README notes. The pipeline runs on a schedule via GitHub Actions, generating files like daily-brief.json, latest-24h.json, and stories-merged.json.

Getting It Running: Fork, Commands, and Configuration

The README gives three entry points. For agent use, it provides a single command: npx skills add LearnPrompt/ai-news-radar -s ai-radar -g. After that, you can ask the agent '今天AI圈有什么？' to get a briefing. For direct web use, you open news.learnprompt.pro or the GitHub Pages mirror. For customization, you fork the repository. The fork guide lists five steps, though the README is truncated; it mentions enabling Actions in the fork. To make it your own, you replace the OPML sources and edit markdown files under personas/ to change the commentary flavors. The data is published to your own GitHub Pages. Configuration relies on environment variables or GitHub Secrets for advanced sources, such as tokens, cookies, private OPML, and email content. The pipeline works without any API key by default, but if you set DEEPSEEK_API_KEY, the pipeline generates real recommendation reasons and enhances short or jargon-heavy titles. Without that key, it falls back to rule-based scoring and keeps original titles. The live site also supports a ?data=<data directory URL> parameter to point the frontend at a different data directory, useful for testing branches.

The Three Personas and the LLM Dependency

A distinctive feature is the persona-based commentary. The README describes three flavors: pragmatic (default), which focuses on practical use for developers; cynic, which debunks marketing hype; and paper-police, which demands evidence from papers, code, or benchmarks. Each persona is a markdown file in personas/ with frontmatter and a system prompt. You can edit a file to change a flavor or create a new one following the format in personas/README.md. The pipeline scores the top 20 daily picks with the default persona, and the top three stories get all three personas. The output goes into data/daily-brief.json and data/top3-personas.json. However, these commentary features require DEEPSEEK_API_KEY. Without it, the pipeline degrades to rule-based scoring, and the web display of the three flavors is already temporarily offline pending redesign, according to the v0.8 release notes. So the persona system is a real feature but not a default one; it is a trade-off between richness and zero-config simplicity.

Story Merging and Source Health: The Core Value

The README emphasizes that the project is not just another news page. Its core logic is the Scout Skill, which evaluates source quality before ingestion. The pipeline classifies sources, skips high-risk ones, and computes source health and AI share. This helps you identify which sources are worth long-term tracking and which are mostly noise. The story merging is another key piece: when two or more sources report the same event, the frontend shows a multi-source N label, and you can expand it to see each source's title, origin, and relative time. This reduces duplicate reading and helps compare wording differences. The data files include stories-merged.json and merge-log.json, the latter for debugging and auditing the merge process. This is a concrete mechanism, not just a UI feature. For content creators, the multi-source signal helps assess credibility: official first-hand, single-source observation, or multiple outlets. This is where the project differentiates itself from a simple RSS reader.

Limitations and Failure Modes

Several limitations are visible from the material. First, the pipeline depends on GitHub Actions scheduling, so the news is not real-time; the README mentions the online site refreshes, but the exact frequency is not stated beyond the action workflow. Second, the recommendation reasons and title enhancement require DEEPSEEK_API_KEY; without it, you get rule-based scores and original titles, which may be cryptic. The README explicitly says that when no real reason is generated, the frontend hides the recommendation block rather than filling it with template text, so the UI can appear sparse. Third, the three-persona web display is currently offline, pending redesign, so you can only see that commentary through the Skill or data files. Fourth, the source classification step can skip high-risk sources, which is a safety measure but also a limitation if you want to include a source that the classifier deems risky. Finally, the README warns that some sources may look active but have low AI relevance, so the AI share metric is essential to filter them; this is a design choice that requires trust in the scoring threshold (score >= 0.3 for the all pool).

Alternatives and the Difference in Approach

A direct alternative is a traditional RSS reader like Feedly or Miniflux. Those tools focus on subscription and reading, but they do not merge duplicate stories across sources, nor do they apply AI relevance scoring or persona commentary. You would still manually triage dozens of items. Another alternative is a custom news aggregator using Python libraries like feedparser and a simple web frontend, but that would require you to build the deduplication and scoring logic yourself. AI News Radar differs by packaging a full pipeline: source classification, fetching, deduplication, relevance scoring, story merging, and persona-based scoring, all in one repository with a static output. The README also mentions sister sites AI MAP and GoodCase, which are separate projects for discovering trends and finding use cases, respectively, but those are not part of this repository. For a user who only needs a simple feed, AI News Radar may be overkill; for someone who wants to curate a high-signal AI news stream, it offers a more opinionated workflow.

Maintenance, Upgrade Cost, and License

The project is under the MIT license, which permits commercial use, modification, and redistribution with attribution. The maintenance cost depends on how you use it. If you just read the live site, you have no maintenance burden. If you fork it, you own the pipeline: you must maintain your OPML sources, update personas if you change them, and ensure GitHub Actions runs correctly. The README notes that the Scout Skill can help agents maintain sources and scraping logic, which could reduce manual work. Upgrades are visible through releases: v0.9 introduced a single-layer information architecture and dual views, while v0.8 added the persona system. The README mentions that the old three-view screenshots are archived in /legacy/ until mid-August 2026, indicating a planned cleanup. The data files are versioned by generation, so you can inspect changes in merge-log.json. There is no explicit upgrade guide in the provided material, so you may need to check the changelog or commit history. The last push is dated 2026-07-14, so the project is actively maintained, but you should verify that the Actions workflow still works in your fork, especially if you use custom sources.

Editorial conclusion

Adopt AI News Radar if you want a self-hosted, customizable AI news feed without API keys or a backend. It suits developers who can fork the repo and tweak OPML sources or personas, and readers who just want the live site. Avoid it if you need real-time updates faster than the GitHub Actions schedule, or if you require deep analytics beyond the provided JSON files. Before adopting, verify that your preferred sources are supported (RSS, OPML, GitHub feeds, static pages, AgentMail) and that you can run the pipeline locally, since the documentation does not detail local setup commands. Also check the current data freshness on the live site, as the last push is from July 2026 and the project may evolve. The core value is the story-merging and persona commentary, but those depend on LLM keys; without them, you get rule-based scoring only.

Official sources

1. Official documentation
2. Official README
3. Project repository
4. Release notes