Launch your own AI hotspot news site: hands-on with AIHOT
Every morning the AI industry publishes a thousand stories and maybe ten that matter. Feed readers hand you the thousand; AIHOT hands you the ten. AIHOT (KKKKhazix/AIHOT, MIT) is a self-hosted website framework that finds its own hot topics and writes its own daily digest: it pulls from six source types, has an LLM pre-screen and then independently score every candidate twice, clusters coverage of the same event, ranks by heat, and publishes a daily edition — the same engine running the live site aihot.news. It went from zero to nearly 6,700 GitHub stars in under two weeks. This walkthrough boots your own instance in about half an hour and shows you how to turn it into a hotspot site for your industry.
- Six source types — RSS, web lists, JSON APIs, X accounts, WeChat public accounts, and content pushed in by your own scripts — split into three tiers (official first-party, official accounts and quasi-official, media and individuals), each with its own admission threshold, and scrape frequency adjusts automatically with output.
- The pipeline: dedup and pre-screen, then double scoring against the same rubric, title rewriting and summary writing (summaries lead with the answer), clustering of same-event coverage, heat ranking, and editions.
- Heat is computed per event, not per article: each independent source counts once inside a 48-hour window, halving after 24 hours.
- Editions: a daily digest compiles at 08:00, a weekly digest on Mondays at 10:00, and a monthly one on the 1st at 10:30 — configurable in site/site.ts under EDITION_TIMES.
- You'll need Docker with Compose, Node.js 24 (for the config generator), and an OpenAI-compatible model API key — the setup script defaults to DeepSeek; Qwen and Zhipu work too. The stack is TypeScript/React Router (SSR), Fastify, PostgreSQL, pg-boss, and Tailwind in one docker compose file; the site opens at localhost:3000 with the admin panel at /admin.
- The repo ships as a template with 18 public overseas AI news sources bundled as demos — but the README warns it deliberately ships without the live site's source list and operating data. Adapting it to an industry mostly means editing site/site.ts, industry/taxonomy.ts + topics.json, industry/sources.json, and industry/prompts/.

The design is the whole pitch: AIHOT treats source curation as a knowledge-management problem. Six source types — RSS, web lists, JSON APIs, X accounts, WeChat public accounts, and content pushed in by your own scripts — are split into three tiers (official first-party, official accounts and quasi-official, media and individuals), each with its own admission threshold, and scrape frequency adjusts automatically with output. A story that survives pre-screening gets scored twice against the same rubric, gets a rewritten title and summary, and is clustered with other coverage of the same event: an LLM decides whether two stories are the same story, a follow-up, or two different things, with a second model able to re-review the decision (GROUP_REVIEW_MODEL). Heat is computed per event, not per article: each independent source counts once inside a 48-hour window, halving after 24 hours, so one outlet publishing ten takes on the same story does not move the needle — what ranks is what a lot of people are actually talking about.
What you’ll need
- Docker with Compose, and Node.js 24 (needed to run the config generator).
- An OpenAI-compatible model API key. The setup script defaults to DeepSeek; Qwen and Zhipu work too, per the README.
- A machine to host it on — the stack is TypeScript/React Router (SSR), Fastify, PostgreSQL, pg-boss, and Tailwind, all in one
docker composefile. - About 30 minutes of patience: the first import’s backlog takes roughly half an hour to work through.
Step 1 — clone and generate your config
git clone https://github.com/KKKKhazix/AIHOT.git myhot
cd myhot
node scripts/init-env.ts --llm-key <your-model-api-key>
init-env writes your .env from a template, configured for DeepSeek by default. If you want Qwen, Zhipu, or another OpenAI-compatible provider, the .env.example file shows exactly which three values to change (LLM_BASE_URL, LLM_MODEL, LLM_EXTRA_JSON). Using a reasoning model that thinks before answering? Also set LLM_REASONING_TOKENS so the reasoning pass does not eat the output budget. The repo ships as a template, so the author suggests clicking Use this template and cloning your own copy if you plan to keep it; a straight clone is fine for a first try.
Step 2 — boot the stack
docker compose up -d --build
Open http://localhost:3000. The admin panel lives at /admin — the password is the ADMIN_PASSWORD init-env put in your .env. Within a couple of minutes the site starts filling with content, processed from the 18 public overseas AI news sources bundled as demos. The README warns you not to mistake these for the real thing: the repository deliberately ships without the live site’s source list and operating data — the demo feeds exist so you can see the pipeline working before you rewire it.
Step 3 — watch the selection pipeline work
In the admin panel you can manage sources, trial a scrape, and inspect the diagnostics of every pipeline step. Here is what each item goes through, in the README’s own order:
- Dedup and pre-screen. A new item is checked for duplicates, then a fast pass discards the obviously uninteresting.
- Double scoring. Anything possibly important is scored twice against the same rubric, independently. Passing the per-tier threshold is what admits it to the curated set.
- Writing. Titles are rewritten, summaries lead with the answer, and a recommendation reason is attached. Foreign-language originals get full translations; the model is explicitly restrained from inventing companies into headlines.
- Clustering. Coverage of the same event is merged: candidate matches are found by vector similarity (or text overlap when no vector service is configured), then the model judges same-story vs. follow-up vs. different story, with a final review pass before writing.
- Heat ranking. Events, not articles, are ranked — rising fast compared to six hours ago gets marked as climbing, newcomers get marked new.
- Editions. A daily digest compiles at 08:00, a weekly digest on Mondays at 10:00, and a monthly one on the 1st at 10:30 — configurable in
site/site.tsunderEDITION_TIMES.

Step 4 — make it your industry’s hotspot
This is the part the author — a designer who rebuilt the codebase with AI half a year after picking it up, and open-sourced it because lawyers, HR people, and finance folks kept asking for their own industry version — actually designed for. Nearly everything lives in two folders, site/ and industry/, and the knowledge-heavy pieces are plain Markdown you can edit without touching code:
site/site.ts— site name, industry vocabulary, edition times, homepage copy, about page, and the categories your public API exposes.industry/taxonomy.tsandindustry/topics.json— categories, tags, and topic pages.industry/sources.json— the sources imported on first boot. Replace the demo feeds with your industry’s.industry/prompts/— every prompt in the pipeline, verbatim. Your industry know-how goes here.industry/selection.ts— the admission thresholds.
The author’s recommended shortcut is to hand the whole repo to a coding agent (Claude Code or Codex) with this instruction:
Please read AGENTS.md and docs/customize.md, and turn this site into a hotspot site for the [YOUR INDUSTRY] industry.
I care about: ... (the sources you want to watch, what counts as hot for you, what doesn't — be specific).
When done, run npm run typecheck, npm test and node scripts/smoke.ts, and tell me what decisions are left for me.
The most valuable thing to get right is the scoring rubric (industry/prompts/selection-score.md) and the thresholds: label a couple hundred items yourself, run scripts/eval-selection.ts to see how well the current settings match your judgment, and iterate. The docs folder has a full calibration guide (docs/: sources, selection and calibration, event grouping, digest evaluation, deployment, architecture).

What you built
Your own hotspot news site: an ingestion engine over six source types, a double-scored LLM curation pipeline with fully public prompts and thresholds, event clustering with cross-source heat ranking, and daily, weekly, and monthly digests — plus the same content exposed for machines: RSS (curated, all, full-text, digests), a public API, an MCP endpoint, /api/v1/agent, an openapi-v1.json spec, and llms.txt. One stack, one site, readable by people and agents alike.
Honest limitations
- It is Chinese-first. The English
README.en.mdexists, but the project’s discussions, issues, and freshest docs are in Chinese. Budget some translation time if you do not read it. - The demo sources are AI-industry only. Until you replace
industry/sources.jsonwith your own industry’s feeds, you have someone else’s demo, not your site. - Every pipeline step spends tokens. Two scores per candidate, clustering judgments, translations — the admin panel has budget fuses for paid services, but price it like a utility, not free software.
- It is a live site’s engine, not a polished framework. The author is explicit: this is the same code running aihot.news, synced over rather than rebuilt as a general framework. Expect rough edges, and do not expect every upstream improvement to land here.
- Some features stay on the live site. The model leaderboard, the Codex reset monitor, themed event chronologies, and the operator tooling never left aihot.news.
- The name and logo are not yours. The code is MIT, but AIHOT’s name and logo are excluded from the license — rebrand before you publish.
Still: the gap between “I follow the AI industry” and “I run the hotspot site my industry reads every morning” used to be a full newsroom. AIHOT closes it to a config file, a rubric, and a Docker command. That is worth an afternoon.