Readr

A self-hosted, single-user RSS reader with full-text article fetching, Reddit comment threads, and a daily AI digest — no cloud, no tracking, no algorithmic interference.

Heads up: This project was vibe-coded — built iteratively with AI assistance, not architected upfront. I use it every day as my primary RSS reader and it works well for me, but it hasn't been battle-tested by a wide audience. Expect rough edges. Issues and PRs are welcome.

Features

• Full-text reading — fetches and stores complete article content server-side via trafilatura
• Topic organization — organize feeds into topics (many-to-many); filter your reading by topic
• Bulk feed import with AI classification — paste a list of URLs (or any text containing URLs) and let the local LLM auto-classify them into topics, with a review step before import
• Inline topic creation — create new topics on the fly when adding a feed, no separate modal needed
• Reddit first-class — Reddit posts include live threaded comment trees, upvotes, and comment counts
• Daily AI digest — local LLM (Ollama + Qwen3.5:9b) summarizes the day's stories per topic at a configurable time
• Article summarization — on-demand per-article summaries via local LLM, cached after first generation
• Named entity extraction — extract people, organizations, places, and topics from articles; view trending entities
• Read tracking & bookmarks — per-article state, persisted in SQLite
• Read later — separate read-later list, distinct from bookmarks
• Tags & automation rules — color-coded tags and fetch-time rules (auto-bookmark, auto-tag, auto-mute)
• Mute filters — drop articles matching patterns (plain text or regex) at fetch time, globally or per-feed
• Smart searches — saved search queries with automatic article matching
• Full-text search — SQLite FTS5, BM25 ranking, no external service needed
• OPML import/export — standard feed list exchange; OPML folders become topics on import
• Article highlights & notes — highlight text passages and attach private notes to any article
• Priority scoring — articles scored based on thumbs-up/down signals per feed
• Mobile-friendly — responsive UI with bottom navigation, accessible over Tailscale from any device
• Self-contained — runs entirely in Docker Compose; all data stays local

Stack

Layer	Technology
Backend	Python 3.12, FastAPI, SQLAlchemy 2 (async), aiosqlite
Frontend	SvelteKit 2, Svelte 5 (runes), TailwindCSS 3
Database	SQLite with FTS5 full-text search
LLM	Ollama (local inference, GPU or CPU)
Feed conversion	RSSHub (Bluesky, YouTube, and 400+ other sources)
Deployment	Docker Compose + nginx reverse proxy

Quick Start

git clone https://github.com/AdrienLF/readr
cd readr

# Pull the Ollama model (first run only)
docker compose up -d ollama
docker exec -it readr-ollama-1 ollama pull qwen3.5:9b

# Start everything (auto-detects GPU)
bash start.sh

Open http://localhost:7755 in your browser.

GPU support

The start.sh script auto-detects nvidia-smi and enables GPU passthrough if available (requires NVIDIA Container Toolkit or Docker Desktop with WSL2 GPU support). To force CPU mode:

docker compose up -d

Adding Feeds

Source	URL format
Any RSS/Atom feed	Direct feed URL
Reddit subreddit	https://www.reddit.com/r/{subreddit}.rss
Bluesky user	http://rsshub:1200/bsky/user/{handle}
YouTube channel	http://rsshub:1200/youtube/user/{username}
GitHub releases	https://github.com/{user}/{repo}/releases.atom

RSSHub supports hundreds of additional sources — see docs.rsshub.app.

Configuration

Settings are available through the UI (/settings) or via environment variables:

Setting	Default	Description
digest_time	07:00	Daily digest generation time (HH:MM)
ollama_model	qwen3.5:9b	Ollama model used for digests
fetch_interval	3600	Feed poll interval in seconds

Environment variables:

Variable	Default
OLLAMA_BASE_URL	http://ollama:11434
RSSHUB_BASE_URL	http://rsshub:1200
DATABASE_URL	sqlite+aiosqlite:////app/data/rss_reader.db

Development

Backend

cd backend
uv sync
uv run uvicorn app.main:app --reload  # dev server on :8000
uv run pytest                          # run all tests

Frontend

cd frontend
npm install
npm run dev    # dev server on :5173
npm test       # unit tests (vitest)

Architecture

readr/
├── backend/
│   └── app/
│       ├── main.py          FastAPI app + lifespan
│       ├── models.py        SQLAlchemy ORM models
│       ├── schemas.py       Pydantic schemas
│       ├── routers/         REST API endpoints
│       └── services/
│           ├── fetcher.py   Feed polling + Reddit normalization
│           ├── extractor.py Full-text extraction (trafilatura)
│           ├── scheduler.py    APScheduler (poll + digest cron)
│           ├── llm.py          Ollama digest generation
│           └── smart_search.py Saved-search matching
└── frontend/
    └── src/
        ├── routes/          SvelteKit pages
        └── lib/
            ├── api.js       Typed fetch wrapper
            ├── stores/      Svelte 5 reactive state
            └── components/  UI components

Security

Readr is designed for local / private network use (e.g. behind Tailscale). It has no authentication layer — anyone who can reach the port can use it. Do not expose it to the public internet.

Protections in place:

• XSS — all rendered HTML from feeds and LLM output is sanitized with DOMPurify
• SSRF — feed URL validation blocks private IPs, loopback, link-local, and internal Docker hostnames
• ReDoS — mute-filter regex patterns are validated against nested quantifiers
• Docker — backend and frontend containers run as non-root users
• Nginx — security headers (X-Frame-Options, X-Content-Type-Options, Referrer-Policy)

Non-Goals

• Multi-user / auth (single-user by design)
• Cloud sync or remote storage
• Twitter/X integration
• Public internet exposure (designed for a private network / Tailscale)

License

MIT