Horizon – Open-Source AI News Aggregator for Your Personal Radar

1. What Is Horizon

Horizon is an MIT-licensed, AI-driven information aggregation system for building a personal AI news radar. An async fetch engine pulls in parallel from RSS, Hacker News, GitHub Trending, Reddit, Telegram, Twitter/X, and OpenBB; LLMs dedupe, score relevance, summarize, and enrich background context; outputs include web dashboards, email digests, webhooks, and MCP—bilingual Chinese/English daily briefs. It frees AI practitioners from noise to focused, high-signal reading.

Image source: Official article

Technical positioning and domain: Personalized aggregation and intelligent summarization—LLM semantic filtering vs passive RSS readers that list everything unread.

Development background: Maintained by Thysrael and the OSS community—born from tracking dozens of sources manually and missing important updates.

Core value: Shifts from passive subscription to AI-curated feeds—dedupe, 0–10 scoring, structured summaries, and background notes in one pipeline.

Technical characteristics: Async multi-source ingestion; pluggable LLMs (OpenRouter, DeepSeek, xAI, etc.); modular outputs (web, email, webhook, MCP); runs with just an LLM API key.

2. Key Features

• Multi-source aggregation: Seven channel types—async parallel fetch; enable/disable per source.

• AI relevance scoring: LLM scores 0–10; buckets high (7–10), medium (4–6), low (0–3) to surface signal.

• Dedup and summarization: Similarity filtering plus structured summaries—key points, data, community highlights—cutting read time.

• Background enrichment: High-score items get extra context—papers, timeline, community takes—for complex topics.

• Multi-format output: Web board, email, webhook, MCP—for humans or agents.

• Flexible filters: Score thresholds, source filters, time windows (24h/3d/7d), keyword match—custom radar rules.

• Bilingual briefs: Daily CN + EN editions—global and domestic AI trends without language gap.

• Modular extensions: Plugin architecture for new sources and sinks—community contributions welcome.

• Signal tiering workflow: After scoring, Horizon buckets items so dashboards emphasize 7–10 "must read" links while still archiving 4–6 for weekly deep dives—reducing FOMO-driven re-reading of low-value reposts common in raw RSS feeds.

3. How to Use

1. Requirements: Python 3.10+, Linux/macOS, Git, uv (recommended) or pip, at least one LLM API key; optional Twitter/Telegram credentials.

2. Clone: git clone https://github.com/Thysrael/Horizon.git && cd Horizon

3. Install deps: uv sync or pip install -r requirements.txt (~5–10 min).

4. Configure: cp config.example.yaml config.yaml—set LLM key/model (e.g. claude-sonnet-4-20250514) and source credentials.

5. Rules: Enable sources; set score cutoffs, time window, keyword filters—start with 2–3 sources for testing.

6. Run: python main.py—first run ~10–20 min; watch logs.

7. Output: HTML under output/pages/; optional email/webhook delivery—tune thresholds from quality feedback.

4. Pros and Cons

5. Comparison with Similar Tools

Selection advice: Technical users wanting ownership and customization—Horizon. Zero-config readers—AIHOT. RSS veterans extending routes—RSSHub (+ optional AI glue).

6. Editor's Review

Horizon systematically applies LLMs across ingest→dedupe→score→summarize→enrich—not just another reader. MCP output fits agent stacks well.

It directly attacks overload for AI builders—score gates focus attention; bilingual briefs help CN researchers track global news.

Effectiveness depends on source curation and model choice; setup is not casual—Docker one-click would help adoption.

— Professional async + LLM design; −0.5 for deploy friction and source dependency; high value for target users.

7. Use Cases

• Daily AI digest: Engineers/PMs scan GitHub/HN/X highlights in ~10 minutes.

• Creator ideation: High-score topics + background for articles and threads.

• Team tech radar: Internal briefs filtered by domain keywords.

• Personal knowledge inbox: Replace raw RSS with scored feeds and time windows.

• Agent knowledge feed: MCP stream for research assistants and consulting bots.

8. FAQ

Q: GPU needed?
A: No local model—external LLM APIs only. 4C8G+ CPU server recommended.

Q: Best LLM?
A: Budget: DeepSeek V3 or Claude Haiku; quality: Claude Sonnet 4 or GPT-4o. Test on OpenRouter free tiers first.

Q: Private deploy / data safety?
A: Fully self-hosted; only LLM requests leave box—use Ollama locally if strict.

Q: Twitter/Telegram access?
A: X dev account (rate limits); Telegram bot token—or disable and use RSS/HN/GitHub only initially.

Q: Tunable scoring?
A: Yes—thresholds, tier cutoffs, keyword allow/deny lists in config.

Q: Multi-user?
A: Single-user core—multiple Docker instances or webhooks to teams; enterprise RBAC needs custom work.

Q: Always-on operation?
A: systemd/Docker + cron daily runs; log rotation; monitor with Uptime Kuma etc.

Q: Can I swap LLM providers without rewriting fetch logic?
A: Yes—config.yaml abstracts provider endpoints and model names. Switching from OpenRouter to DeepSeek or a local Ollama endpoint typically requires only key and model string changes, then a short validation run to recalibrate score thresholds because different models rank relevance differently.

9. Project Links

• GitHub: https://github.com/Thysrael/Horizon
• Website: https://www.horizon1123.top/