A website framework that finds the news and writes the daily briefing by itself.
Swap in your own sources and your own know-how for what matters, and it becomes the news site for your field.
## What this is
[AIHOT](https://aihot.news) is an AI news site I built. Every day it collects material from a set of sources. A language model screens every item, then scores the promising ones twice, independently, to pick what is really worth reading, and writes a headline and a summary for each. Reports of the same story from different sources are clustered into one event, and events are ranked by how many sources are talking about them. Every morning it publishes a daily briefing.
This repository is its engine and framework: the website, the admin, the selection pipeline, the clustering and heat algorithms, and **the full text of every prompt and every selection threshold**.
## Why open source
Over the past half year, many friends working in law, HR, finance and precious metals have asked me whether I could build one for their field.
I can't. I don't know your field. I don't know which sources are useful, or what kind of news counts as news to you.
But you do.
Since I can't do it for everyone, I'm handing the spark to all of you.
## Before you start
- **The interface and everything the site writes are in Chinese.** The prompts in [`industry/prompts/`](industry/prompts/) set the output language; the interface text is in [`site/site.ts`](site/site.ts) and in the code. To run it in another language, ask your agent to translate those (see [Make it your field](#make-it-your-field)). The other documents in this repository are in Chinese too; coding agents read them without trouble.
- **I'm not a professional developer.** I started out as a designer, and half a year ago I could barely read code. I rewrote this codebase together with AI. It is much cleaner than before, but there are surely parts that aren't written well. Issues are welcome. I may not reply quickly, and I apologise in advance.
- **This is AIHOT's engine.** It is the same engine code that runs AIHOT in production, exported directly from it, not a carefully polished general-purpose framework. I'll try to bring AIHOT's updates over, but I can't promise every one. Features that only make sense for AI, such as the model leaderboard, the Codex reset monitor and the milestone timelines on topic pages, stay on AIHOT, along with AIHOT's own operations tools.
- **AIHOT's source list and operating data are not included.** The repository ships 18 public AI news sources as a demo, enough to run it and see what it does. The real sources should be your own field's.
- **Please don't use AIHOT's name or logo.** Put your own name on it, and it's your site.
## How it works
An item comes in from a source, is checked for duplicates, and pre-screened. Items that might matter are scored twice, independently, get a headline and a summary, are clustered with other reports into an event, and count toward its heat. X posts with no text or only links keep their original links and media without a writing or translation model call. An item enters the selection only if its score clears the threshold and it doesn't repeat news already selected. The daily briefing is compiled from the day's top stories by rule; weekly and monthly reports are compiled from the dailies. Every step's prompt is in [`industry/prompts/`](industry/prompts/), so changing the standard doesn't mean changing code. See [Selection and calibration](docs/selection.md).
### Clustering and heat
The same thing happens once: the official site posts it, ten outlets repeat it, X argues about it all day. Readers only need to see it once. AIHOT clusters these into one **event**. It looks for candidates from the past two weeks by embedding the headline and summary (or, without an embedding service, by text overlap), then asks a model whether they are the same story, a follow-up, or two different stories. When in doubt it merges, and before writing, a model reviews the merge again (the review can use a different provider; set `GROUP_REVIEW_MODEL`).
**Heat** is computed per event, not per article: within 48 hours, each independent source counts once, halving every 24 hours. Repeated fetches don't add up, and one outlet posting ten articles still counts once, so what ranks at the top is what many people are really talking about.
### Speed
## What you get
| | |
|---|---|
| **Six kinds of sources** | RSS, web listing pages, JSON APIs, X accounts, WeChat Official Accounts, and content your own scripts push in. Sources come in three tiers (official first-hand; official and semi-official accounts; media and individuals), each with its own selection threshold. Fetch frequency adjusts to how often a source publishes |
| **Selection** | A pre-screen, then two independent scores on the same rubric, then a threshold set by the source's tier. One piece of news takes one slot; the same news reworded doesn't get in. All prompts and thresholds are public and editable; calibrate them in SelectBench with samples you labelled yourself |
| **Writing** | Headlines, answer-first summaries, reasons to read, full translation of foreign-language text; categories, tags and news facts extracted separately; guards against the model putting companies the original never mentioned into a headline |
| **Clustering** | Reports of the same story from different sources become one event, with follow-ups under the same event and an overview on the event page. Follow-ups and the report timeline switch together between newest first and oldest first. Manual corrections to clustering are never overwritten |
| **Trending** | Heat per event: the more independent sources, the higher, and discussion on X counts too. Compared with six hours earlier, fast risers are marked rising and newcomers are marked new |
| **Daily, weekly and monthly reports** | A daily briefing every day (08:00 by default), compiled from the day's top stories by rule: one entry per story, a story already covered returns only when it has news, and no model is called. A weekly report every Monday and a monthly report on the 1st (10:00 and 10:30 by default), compiled from the dailies; the model writes only the overview and section introductions. Change the times in `EDITION_TIMES` in `site/site.ts` |
| **Topics and search** | Topic pages of three kinds: companies, directions and content formats; search over headlines and summaries, and full-text relevance search |
| **Reading preferences** | Light and dark appearance, with the site's default font or the browser's default font. Saved items and reading preferences stay on the device and can be exported and imported as a backup |
| **For agents** | RSS (selection, everything, full text, daily, weekly, monthly), a public API, MCP, Agent Markdown and `llms.txt`: the same content for people and for agents |
| **Admin** | Source management and trial fetches, content diagnostics and manual score corrections, selection evaluation, a separate model for each step, budget breakers for paid services, run history and alerts |
## A quick look
Screenshots from a local site running on the demo sources, under the default name MyHOT.
## Run it
To create your own independent site, click [Use this template](https://github.com/KKKKhazix/AIHOT/generate) and clone the repository it creates. To keep merging upstream updates or to contribute, fork it instead. The commands below are for trying it out directly.
You need [Docker](https://docs.docker.com/get-started/get-docker/), [Node.js 24](https://nodejs.org/en/download) (to run `init-env`, which generates the configuration), and an API key for any OpenAI-compatible model API (DeepSeek, Qwen, Zhipu GLM and so on).
`init-env` configures DeepSeek by default. For another provider, change `LLM_BASE_URL`, `LLM_MODEL` and `LLM_EXTRA_JSON` in `.env`, following the examples in `.env.example`. With a reasoning model (one that thinks before answering), also set `LLM_REASONING_TOKENS` to leave output room for the reasoning.
```bash
git clone https://github.com/KKKKhazix/AIHOT.git myhot
cd myhot
node scripts/init-env.ts --llm-key
docker compose up -d --build
```
Open . The admin is at `/admin`, and the admin password is `ADMIN_PASSWORD` in `.env`. Content starts to appear after a minute or two, and the first import takes about half an hour to process.
For machines without Node, servers in mainland China, domains and HTTPS, or running directly on Linux or macOS without Docker (WSL2 on Windows), see [Deployment](docs/deploy.md).
Once the site is up, `/agent` lets you copy how to connect over MCP, RSS or the API; agents that can only read web pages start from `/api/v1/agent`; the API description is at `/openapi-v1.json`.
## Make it your field
The easiest way: open your coding agent (Claude Code or Codex both work), give it this repository, and say:
```text
Please read AGENTS.md and docs/customize.md, and turn this site into a news site for the "XX" field.
What I care about: … (the sources you want to watch, what news you think matters and what doesn't; the more specific the better).
If the site should be in English (or another language), also translate the interface text and rewrite the prompts in industry/prompts/ to write in that language.
When you're done, run npm run typecheck, npm test and node scripts/smoke.ts, and tell me which decisions are still mine to make.
```
Nearly everything to change is in the [`site/`](site/) and [`industry/`](industry/) folders; the code itself mostly stays as it is:
| File | What to change |
|---|---|
| `site/site.ts` | Site name, field terms, publishing times, home page copy, about page, the wording on items and reports, the categories of the public interfaces |
| `industry/taxonomy.ts`, `industry/topics.json` | Categories, tags, topics |
| `industry/sources.json` | The sources imported on first start |
| `industry/prompts/` | The selection standard and the writing requirements. **Your field's know-how goes here** |
| `industry/selection.ts` | Selection thresholds |
| `site/models.ts` | Which model each step uses by default (optional: otherwise every step uses the one in `.env`) |
| `site/brand/`, `site/pages/`, `site/public/`, `site/changelog.json` | Icons and logo, terms of use and privacy notice, files served from the site root such as `robots.txt`, the changelog |
| `modules/` | Features the framework doesn't have and only your site needs, as modules; see “模块” in [Architecture](docs/architecture.md#模块) |
What is most worth your time is the scoring rubric (`industry/prompts/selection-score.md`) and the thresholds: take one or two hundred items you labelled yourself, run them through `scripts/eval-selection.ts`, see how accurately it selects, and go back and adjust. How to do it is in [Selection and calibration](docs/selection.md).
## Documentation
The documents below are in Chinese.
| Document | Contents |
|---|---|
| [Make it your field](docs/customize.md) | Site name, categories, topics, sources, prompts, thresholds, models, brand, step by step |
| [Sources](docs/sources.md) | How to configure the six kinds of sources, tiers and full text, fetch frequency, older articles and archives, a fixed starting point with `publishedAfter`, the external push interface |
| [Selection and calibration](docs/selection.md) | How an item becomes a selection and goes into the daily, weekly and monthly reports, and how to calibrate with your own samples |
| [Event grouping and relation evaluation](docs/grouping.md) | How relations between events are judged, and how to evaluate with your own labelled pairs |
| [Overview evaluation](docs/story-digest-evaluation.md) | Before changing the event overview prompt, how to compare versions side by side on the same events |
| [Deployment](docs/deploy.md) | Docker, domains and HTTPS, mainland China, updates, backups, what it costs, and running on Linux or macOS (WSL2 on Windows) without Docker |
| [Architecture](docs/architecture.md) | Three processes, the rules that don't change, directories, modules, database migrations, public outputs, tests |
Stack: Node.js 24 · TypeScript · React Router (server-side rendering) · Fastify · PostgreSQL · pg-boss · Tailwind CSS · Docker Compose.
## Community and contributing
Questions about deploying and using it go to [Q&A](https://github.com/KKKKhazix/AIHOT/discussions/categories/q-a), new ideas to [Ideas](https://github.com/KKKKhazix/AIHOT/discussions/categories/ideas), and you're welcome to share the news site you built for your field in [Show and tell](https://github.com/KKKKhazix/AIHOT/discussions/categories/show-and-tell). Writing in English is fine.
For a bug or a concrete feature request, [open an issue](https://github.com/KKKKhazix/AIHOT/issues/new/choose). Before changing code, read the [contributing guide](CONTRIBUTING.md); report security vulnerabilities through the [private reporting channel](SECURITY.md).
## Finally
AIHOT was once just a tiny, tiny thought on one of countless late nights.
I don't know what it will be turned into, or how far it will go. But that may be the most romantic thing about open source.
The rest of the road is yours.
— 数字生命卡兹克 (Khazix)
## License
The code is under the [MIT License](LICENSE). The AIHOT name and logo are not covered by the license. The fonts have their own licenses; see [NOTICE](NOTICE).
