Newsgetter

A personal news aggregator. It collects articles from RSS feeds (and similar sources), scores them with an LLM according to your interests, and surfaces the ones worth your attention.

Features

• RSS/Atom feed collection with deduplication
• LLM-based relevance scoring (0–100) with a short justification shown next to every article
• Topic tabs with categories assigned by the LLM
• A "Rejected" view of the last 7 days, so false negatives can be restored
• Periodic source checking (default: every 4 hours) with scoring right after every fetch
• Automatic retention cleanup (rejected after 7 days, read after 21 — both configurable)
• Application log with rotation and an in-GUI viewer
• Single SQLite database, no external services required

Quick start

The image is published on Docker Hub as qbsoon/newsgetter:latest:

docker compose up -d

The web interface is available at http://localhost:8000.

To build the image from source instead:

docker build -t qbsoon/newsgetter:latest .

The container starts as root and immediately drops privileges to whoever can write the mounted data directory, so any bind-mount just works — no host-side chown or compose changes needed:

• Mount owned by a regular user → the app runs as that user.
• Root-owned mount (e.g. TrueNAS datasets, root:root with modify rights for the apps user via ACLs) → the app probes well-known unprivileged identities (apps 568:568, user 1000:1000, the mount's group) and runs as the first one that can write; only falls back to root if none can.
• Named volumes work out of the box.

You can force an identity with the PUID/PGID environment variables.

Configuration

All persistent state (database, logs, settings, feed list) lives in the data directory: data/ next to the application, in Docker /srv/data (mount it as a volume). Settings can be changed in the GUI, by editing settings.yaml there, or with environment variables, which always win over the file.

YAML key	Environment variable	Default	Description
host	NEWSGETTER_HOST	0.0.0.0	Bind address
port	NEWSGETTER_PORT	8000	Listen port
refresh_interval_hours	NEWSGETTER_REFRESH_HOURS	4	How often sources are checked
rejected_retention_days	NEWSGETTER_REJECTED_DAYS	7	Days to keep rejected articles
read_retention_days	NEWSGETTER_READ_DAYS	21	Days to keep read, non-rejected articles
llm_provider	NEWSGETTER_LLM_PROVIDER	chutes	chutes or opencode
llm_model	NEWSGETTER_LLM_MODEL	—	Model name used for scoring
llm_api_key	NEWSGETTER_API_KEY	—	API key for the LLM provider
llm_batch_size	NEWSGETTER_BATCH_SIZE	15	Articles per scoring request
llm_timeout	NEWSGETTER_LLM_TIMEOUT	300	Seconds to wait for one LLM request
llm_no_think	NEWSGETTER_LLM_NO_THINK	false	Append Qwen3's /no_think to prompts
score_threshold	NEWSGETTER_SCORE_THRESHOLD	40	Scores below this are rejected by the LLM
profile_batch_minutes	NEWSGETTER_PROFILE_BATCH_MINUTES	45	Minutes between batched profile updates
sources_file	NEWSGETTER_SOURCES_FILE	data/sources.yaml	Path to the feed list

llm_api_keys (per-provider keys), provider_base_urls and llm_extra_body (extra request fields, e.g. to disable a reasoning model's thinking mode) are YAML-only.

The path to the YAML file can be changed with NEWSGETTER_CONFIG.

Adding feeds

The feed list is stored as sources.yaml in the data directory, so it survives container recreation. Add sources in the GUI or edit the file by hand:

sources:
  - name: Ars Technica
    url: https://feeds.arstechnica.com/arstechnica/index
  - name: Nature
    url: https://www.nature.com/nature.rss

Any RSS/Atom feed works this way.

Entries are deduplicated by URL (tracking parameters are ignored), so re-checking a feed is safe.

YouTube channels

Channels are tracked through their official RSS feed. Use type: youtube with the channel id (or a playlist id):

sources:
  - name: Veritasium
    type: youtube
    channel_id: UCHnyfMqiRRG1u-2MsSQLbXA

Only the title, the channel name and a short description are stored.

LLM scoring

After every refresh, new articles are scored in batches (score 0–100, a one-sentence justification, a category) by the configured provider. Scores at or above score_threshold are accepted; lower ones are rejected by the LLM. Each article is scored exactly once. A reader interest profile is kept in the database and updated by the LLM from user actions (read/reject/ restore); signals are batched and applied once per profile_batch_minutes in a single request. The profile feeds the scoring prompt and can also be edited manually.

Token usage is recorded per request. Current-month usage per provider/model/purpose is available at GET /api/usage.

Maintenance

Sources are checked automatically every refresh_interval_hours (changeable in the GUI — applies immediately). After every cycle new articles are scored and retention cleanup runs: rejected articles older than rejected_retention_days and read, non-rejected articles older than read_retention_days are deleted; unread accepted articles are kept.

Application logs are written to logs/newsgetter.log in the data directory (with rotation) and can be browsed in the GUI under "Logs". LLM token usage is shown on the Settings page and at GET /api/usage.

Development

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
hypercorn --reload --bind 127.0.0.1:8000 "app:create_app()"

Project layout

app/          Application code (Quart)
config/       Configuration files
static/       Stylesheets and other static assets
templates/    Jinja2 templates
data/         Runtime data (database) — created automatically