Horizon is a self-hosted "news radar": you list the places you read, describe what you care about, and it writes a ranked daily briefing for you. The project, by GitHub user Thysrael, was listed on GitHub's weekly trending page for week 45 of 2026. The repository now shows roughly 12.3k stars, 1.8k forks and 301 commits under the MIT license.
This guide explains what it does, how to run it, what it will cost you, and where it can mislead. It is a tool explainer based on the README and repo contents, not a hands-on benchmark. We did not run it for this post.
TL;DR: questions people ask first #
| Question | Answer |
|---|---|
| What does it do? | Collects items, scores them with an LLM, dedupes, enriches with context and discussion, and writes a briefing |
| Which sources? | Hacker News, RSS/Atom, Reddit, Telegram, X, GitHub, OpenBB, OSS Insight, GDELT, Google News |
| Which languages? | English and Chinese briefings (README also in Japanese) |
| License? | MIT |
| Do I need an API key? | Yes for most providers; ChatGPT subscribers can sign in instead, per the README |
| Where does output go? | Markdown files, GitHub Pages, email, webhooks (Feishu/Lark, DingTalk, Slack, Discord), WeChat |
| Hosted version? | A live demo site exists; the main path is self-hosting |
| Is it on PyPI? | Not yet; listed as planned |
What is Horizon, in plain words? #
The README's pitch is that good reads are scattered across feeds, forums and timelines, and attention is limited. Horizon collects items, filters and deduplicates them, then brings the worthwhile ones to you with context and community discussion.
The pipeline described in the README runs in this order:
- Collect from the sources you configure.
- Route each item to one profile.
- Analyze, filter and deduplicate with an LLM, before spending tokens on enrichment.
- Enrich the survivors with web-researched background and community comments.
- Generate a briefing grouped by profile, in English and Chinese.
- Deliver by file, Pages site, email, webhook or WeChat.
Doing filtering before enrichment is a sensible cost choice. The expensive step, researching background, only runs on items that already cleared your score threshold.
Profiles: the idea that sets it apart #
A profile is "a reusable set of editorial rules: what belongs, what's worth keeping, and what to write." Three built-in examples ship with the repo:
| Profile | For | Output |
|---|---|---|
| tech-news | Technology news | Events and context, with impact and discussion when useful |
| tech-blog | Engineering deep dives | Background, solution, and practical takeaways |
| ai-creator | AI creator material | A summary, with timely hooks and content angles |
You can assign a profile to a source, set it to "auto" so the AI picks, or restrict matching to a list such as tech-news and finance-news. Thresholds and topic deduplication live in processing.profile_settings, not in the profile files. In the README's example, tech-news has a threshold of 7.0.
This matters because a single scoring prompt cannot serve both a news story ("why it matters") and an engineering post ("what can I use"). Separate profiles let you tune each one.
How to set it up #
The README gives two paths. This is the local one:
git clone https://github.com/Thysrael/Horizon.git
cd Horizon
uv sync
uv run horizon-wizard
uv run horizon
The wizard asks about your interests, with examples like "LLM inference" or "web security", and writes data/config.json. If you prefer manual setup, copy .env.example to .env and data/config.example.json to data/config.json, then add your API key. The config field api_key_env must hold the name of an environment variable, not the key itself.
A minimal source entry looks like this in the README: an RSS source named for a blogger, a URL, and a profile. Useful flags are --hours (default 24), --data-dir, --config and --log-level.
For Docker: docker compose run --rm horizon. The README warns that the optional twitter extra needs a Playwright browser and system packages that the current Dockerfile does not install.
Keep one topic from crowding out the rest
An optional digest block caps the briefing length (for example, 20 items) and gives each category group a limit, such as 5 for AI and 5 for finance. Group limits apply after profile filtering and before enrichment. That is a good defense against a busy week of one model release flooding everything else.
Run it on a schedule #
The repo includes a daily GitHub Actions workflow template. It is disabled in the repository, so you configure it for your deployment and rename it to daily-summary.yml. With GitHub Pages, generated Markdown is copied into docs/ and published as a daily-updated site. Email delivery handles subscribe and unsubscribe through SMTP and IMAP, and webhook notifications can post results or failures to team chat.
Using it with an AI assistant and MCP #
Horizon ships an agent skill so an assistant can configure sources, customize profiles and run the pipeline:
npx skills add Thysrael/Horizon --skill horizon -g
The command needs Node.js 22.20 or newer. The README also documents an MCP server that exposes pipeline stages as tools. If MCP is new to you, start with our MCP explainer; for the skill format, see what agent skills are. Our list of the top open and closed-source MCP servers shows where Horizon's approach sits among them.
How does it compare with other tools? #
Horizon is a DIY pipeline, not a hosted service. Here is how it differs from related coverage on explainx.ai:
- Cloudflare Zero Wire is a hosted briefing hub for critical CVEs. Horizon lets you build your own feed on any topic.
- claude-mem adds memory to a coding agent. Horizon does not touch your coding workflow; it filters what you read.
- CowAgent is a full agent harness with channels. Horizon is narrower: collect, score, brief.
- Fast-growing repos like MoneyPrinterTurbo show how sponsor lists appear in READMEs; read them with care.
What it will cost you #
The software is free. The costs are:
- LLM calls: each collected item may be scored, and survivors are enriched. More sources and a lower threshold mean more calls. The README does not give a per-day cost figure, so measure your own with a small --hours value first.
- X access: the README says Twitter and X timelines and searches go through Apify, a paid third-party scraper.
- Hosting: a free GitHub Actions run may be enough for a small feed, though we have not tested limits.
If you have a ChatGPT plan, the README says eligible subscribers can pick the openai-chatgpt provider and use Sign in with ChatGPT instead of an API key. Check OpenAI's current terms for automated use of a consumer plan before relying on that for a scheduled job.
Honest limits and risks #
- LLM scoring is a judgment call. A threshold of 7.0 is only as good as your rubric. Treat the briefing as a reading filter, not a fact check.
- Scraping dependencies. Reddit, Telegram and X sources depend on access that platforms can change.
- Sponsors in the README. The Supported By section lists commercial API resellers, including one that advertises large discount multipliers on GPT and Claude access, and a cache-hit refund promise. We have not verified any of those claims. Reselling model access at steep discounts can raise questions about terms of service and data handling, so do not send private data through a provider you have not vetted.
- Data flow. Every item and comment you collect goes to your chosen LLM. Do not point private feeds at an untrusted endpoint.
- Security reporting. The repo includes aSECURITY.md and anMIT LICENSE file; read both before deploying it for a team.
- Early packaging. The roadmap lists GitHub releases and PyPI publishing as still planned, so expect to pin a commit.
Who is it for? #
It suits readers who follow many sources and want a morning digest with a point of view: researchers, founders, and developer-relations people. It is less useful if you only read two newsletters. It also serves teams: webhook delivery to Slack or Feishu makes it a shared radar.
If you track AI news closely, the profile idea is the part worth copying even if you never install Horizon. Write down, in plain language, what belongs in your feed, what clears the bar, and what format you want. Then hand that to any tool.
Try this: a starter plan for a first week #
- Install and run the wizard with three interests only.
- Add five RSS feeds you already trust and Hacker News.
- Run with --hours 24 and read the output critically. Note what it got wrong.
- Raise or lower the threshold by one point and rerun.
- Add a category cap to the digest block.
- Only then add Reddit, Telegram or X, and watch your API spend.
Related reading #
Details are from the Horizon README and repository page as of November 4, 2026. Star counts and features change quickly.