OpenPaper
A personalized newspaper that knows only you will ever read it.
OpenPaper is a Claude Code plugin that turns any collection of news sources into a single-reader daily newspaper. You point it at URLs, RSS feeds, or topics; it writes Python fetchers that scrape them deterministically, curates the articles to your taste, and renders the result as a broadsheet-style HTML page — complete with blackletter masthead, multi-column grid, halftone images, and animated ink-draw headlines.
There is no server, no account, no feed algorithm optimizing for engagement. Just your sources, your preferences, and a newspaper laid out for one.
How it works
Three stages, each with a clear boundary:
- Ingest — Add any news source. An agent analyzes the site and writes a standalone Python fetcher. Fetchers import shared utilities from a base module and run without AI on every subsequent invocation — deterministic, fast, auditable.
- Curate — Claude filters, ranks, and selects articles based on your preference profile. Give feedback ("more climate", "fewer opinion pieces") and the profile sharpens over time.
- Present — Claude assigns articles to layout slots and renders the edition using a Jinja2 broadsheet template.
Ingestion pipeline
Under the hood, the Ingest stage runs in two phases:
- Listing — Each fetcher runs in
--listing-only mode, outputting article metadata (title, URL, date) without fetching content. RSS and API fetchers skip Playwright entirely in this phase. All fetchers run in parallel.
- Content — The pipeline deduplicates listings against
seen.txt, then batch-fetches full article text for new articles only, using a single shared Playwright browser. Paywall detection, content extraction, and HTML caching are centralized.
This architecture (inspired by TriOnyx newsagg) avoids wasted work: content is only fetched for articles that pass dedup, and the browser only launches when there's something new to fetch.
Data layout
Your data lives in .openpaper/ inside your project:
.openpaper/
├── sources/ # one .py file per source (+ _base.py shared module)
│ ├── _base.py # shared utilities (deployed by fetch_all.py)
│ ├── hackernews.py
│ ├── bbc.py
│ ├── nrk.py
│ └── ...
├── preferences.md # your interests + feedback history
├── incoming/ # articles awaiting curation
├── saved/ # archived articles from past editions
├── editions/ # rendered HTML newspapers
├── cache/ # per-source HTML + listing caches
└── seen.txt # centralized dedup log (URLs)
Install
Requires Claude Code and uv. Playwright's Chromium must be installed once:
uv run playwright install chromium
Standalone (own project)
Clone the repo and start Claude Code inside it:
git clone https://github.com/falense/openpaper.git
cd openpaper
claude
Plugin (inside another project)
Install OpenPaper as a Claude Code plugin to use it from within any project:
claude plugin add https://github.com/falense/openpaper.git
The /openpaper skill and all commands will be available in your project. Your data (sources, editions, preferences) lives in .openpaper/ inside your working directory — not inside the plugin.
Usage
/openpaper
Or just say:
- "Make my paper"
- "Add nrk.no as a news source"
- "Show me the news"
- "I want more climate articles"
On first run, OpenPaper walks you through adding sources, setting preferences, and generating your first edition.
Fetcher types
| Type | Listing method | When to use | Example sources |
|---|
| RSS | feedparser + urllib | Site publishes an RSS/Atom feed | bbc, nrk, vg, openai, importai, thegradient |
| Playwright | Headless Chromium | JS-rendered listing page, no feed available | anthropic, deepmind, kode24 |
| API | httpx | Documented JSON API | hackernews |
Each fetcher is a standalone PEP 723 Python script — uv run handles dependencies automatically. See the fetcher guide for the full contract.
Template
A single responsive broadsheet that scales from 375px mobile to 1960px desktop spread. Features UnifrakturCook blackletter masthead, paper grain, halftone image effects, and animated ink-draw headlines.
License
MIT
