Meetapedia
A directory of local community groups — running clubs, choirs, board-game nights,
climbing crews, book circles — collected from the open web, city by city.
Two sites, one project:
- meetapedia.com — the international edition, cities worldwide, English (plus other languages).
- közösségek.com — the Hungarian edition: the same directory in Hungarian, for Hungarian cities.
Why it exists
Finding a local group to join is oddly hard. The information exists — on club pages,
municipal sites, parish newsletters, sports association listings — but it is scattered
across thousands of pages nobody thinks to search, in the local language, often without
so much as a meeting time on the front page.
Meetapedia reads those pages so you don't have to. For every city × interest pair it
searches the open web, downloads what it finds, and uses a language model to pull out the
few things that actually matter: what the group is, where and when it meets, and how to
reach it. Groups you can't actually join — commercial classes, one-off events, dead pages —
are filtered out.
At the time of writing it tracks 774 cities across 63 countries and 36 interest topics.
Who makes it
A hobby project by P. Tóth András, built
and run in the open. Nothing here is a company, a product, or a business — it is one
person's answer to "why is this so hard to look up?"
Everything is public: the pipeline, the extraction prompts, the operational history, and
the engineering wiki. If a listing on the site looks wrong, you can read the exact code
path that produced it — every public page links back to the source it was extracted from.
How it works
For each (city, topic) pair:
- Search queries are built in the city's own language and sent to a search API.
- Result pages are fetched and stripped to clean text (social media domains are skipped).
- A language model extracts structured records — communities, venues, and people.
- Records are deduplicated, quality-gated, and stored in SQLite.
- A FastAPI app serves both domains from a single container.
Everything is aggressively cached and fingerprinted, because the whole thing runs on a
hobby budget: pages are downloaded once, extractions are re-run only when the prompt or
model changes, and the expensive LLM work is scheduled inside the provider's off-peak
discount window.
Documentation
The real documentation is the engineering wiki — an LLM-maintained
knowledge base of how the system actually behaves, including the parts that went wrong.
Good starting points:
CLAUDE.md in the repo root is the working brief for coding agents (and a decent
orientation for humans): commands, architecture, and the patterns that are easy to get
wrong. AGENTS.md is a generated copy of it.
Running it yourself
pip install -e ".[dev]"
PYTHONPATH=. pytest # tests
.venv/bin/ruff check scraper/ # lint
The deployment is Docker on Coolify (Hetzner); only /app/data (SQLite) and
/app/config (YAML) are persisted. ADMIN_PASSWORD is required; DEEPSEEK_API_KEY,
DATAFORSEO_LOGIN/DATAFORSEO_PASSWORD, RESEND_API_KEY and the GA4 credentials are
optional — missing keys degrade to a no-op rather than an error. See
deployment for the full list.
License
MIT — do what you like with the code. The community data itself is extracted
from public web pages and belongs to the communities that published it.
Data quality
Every record is machine-extracted from a public web page. Information can be incomplete,
outdated, or plain wrong; nothing is verified by hand. Always check with the community
itself before showing up. Each public page carries a link to its source, a "report this"
button, and an edit-suggestion form — corrections are reviewed and applied.
