News Digest
AI-powered news aggregation and summarization platform with a newspaper-inspired editorial UI. Fetches RSS feeds, summarizes them using LLM APIs, and presents everything in a clean, multi-column layout with real-time widgets.
Features
Core News
- RSS feed aggregation with AI-powered summarization
- Provider fallback system (Groq / OpenRouter) with automatic failover
- Summary history with date-based archive navigation
- Article-level caching to minimize redundant API calls
- Sentiment analysis and topic extraction per article section
- Feed auto-discovery — paste any URL to find RSS feeds
- Category-level custom prompts and language settings for LLM output
Interactive
- Chat with AI about any summary ("Ask about this news")
- Morning briefing — cross-category daily digest
- Model selection — switch between Llama and MiniMax2.7
- Send summaries to Telegram with one click
- Pull-to-refresh on mobile
Widgets
- Job board — aggregated from 12 sources (incl. direct Greenhouse/Lever/Ashby/Workable boards) with AI filtering driven by a configurable
JOB_PROFILE
- Upcoming movies & TV shows (TMDB) with detail modals
- Hacker News top posts
- Weather forecast (Open-Meteo)
- Cryptocurrency prices (CoinGecko)
- RON exchange rates (Frankfurter)
- Trending topics from your feeds
- PBS Headlines
MindGames — Cognitive Resilience Dashboard
Based on the "Mental Antibody" implementation spec and the "Engineering Cognitive Resilience" research paper. 57 features across 5 tabs:
Overview — Dashboard stats, quick actions, recent activity
Training
- Inoculation Lab (Van der Linden) — Build "mental antibodies" by identifying manipulation tactics in AI-generated headlines. Includes Detective Mode + CDO Mode with level progression, score tracking, and
useOptimistic for instant feedback
- Pattern Recognition Tests — Interactive exercises demonstrating apophenia (seeing patterns that aren't there) using random data: coincidence, scatter, sequence, and correlation tests
Analysis
- Cognitive Forensic Engine (Grimes/Ariely) — Deconstruct text into logical and psychological components. Fallacy detection with 14 taxonomy types, emotional intensity mapping, Funnel of Misbelief visualization, SSE streaming, history viewer
- Study Stress-Tester — Evaluate news headlines about studies: sample size, control groups, conflicts of interest, peer review status
- Compare Coverage (Ground News style) — Multi-column Left/Center/Right outlet comparison with framing analysis and narrative divergence scoring
- News Spectrum — Side-by-side outlet credibility ratings and bias spectrum visualization
Reflection
- Scientist's Sandbox (Adam Grant) — ADEPT multi-agent debate (Skeptic, Institutionalist, Moralist), pre/post confidence tracking, thinking mode detection (Preacher/Prosecutor/Politician/Scientist), rethinking journal with trend charts
- Bridge Builder (Monica Guzman) — SOS Audit (Sorting/Othering/Siloing), "How" bridge-building questions, Schwartz Values Quiz, Information Diet echo-chamber diagram
- Stress & Bias Diagnostic — Pre-engagement emotional state check-in with motivated reasoning risk score
Reference
- Prompt Library — Searchable templates for bias detection, perspective taking, fallacy search, steelmanning, source evaluation
- Narrative Map — Misinformation spread visualization across platforms with timeline slider and mutation tracking
- Disinfo Influencer Map — Gateway (health/wellness) → Bridge (influencers) → Conspiracy Core funnel visualization
- Daily Quiz — Daily technique-guessing quiz bridging cognitive and bias-radar modules
Bias Radar — 5-tab slide-over panel used across the app (from article views and homepage): technique picker, compare, decode, steelman, timeline, diet report, gut check
Design
- 4 newspaper color themes (Classic, Broadsheet, Evening, Morning)
- Three-column layout: left sidebar, main content, right sidebar
- Editorial typography: Playfair Display, Lora, Inter, Source Sans 3
- LLM usage statistics with live provider quota tracking
- 5-column newspaper grid homepage with article images
- Server-synced theme preference
Tech Stack
| Layer | Technology |
|---|
| Frontend | React 19, TypeScript, Tailwind CSS 4, Vite, ShadCN UI, Recharts |
| Backend | Express 5, better-sqlite3 (WAL mode), rss-parser |
| AI | Groq (openai/gpt-oss-20b), OpenRouter (minimax/minimax-m2.7) |
| APIs | TMDB, CoinGecko, Open-Meteo, Frankfurter, Hacker News, PBS |
Getting Started
Prerequisites
Installation
# Clone the repository
git clone https://github.com/elodlukacs/news-digest.git
cd news-digest
# Install server dependencies
cd server
npm install
cp .env.example .env
# Edit .env with your API keys (see Configuration below)
# Install client dependencies
cd ../client
npm install
Configuration
Edit server/.env with your API keys:
# Required — at least one LLM provider
GROQ_API_KEY= # https://console.groq.com
OPENROUTER_API_KEY= # https://openrouter.ai
# Optional
TELEGRAM_BOT_TOKEN= # BotFather on Telegram
TELEGRAM_CHAT_ID= # Your chat ID
TMDB_API_KEY= # https://www.themoviedb.org/settings/api
DB_PATH= # SQLite path (default: ./newsreader.db)
PORT= # Server port (default: 3001)
# Access control — see "Securing a deployed backend" below
API_TOKEN= # Bearer token required on every endpoint except /api/health
ALLOWED_ORIGINS= # Comma-separated browser origins allowed to call the API
# Timeouts
LLM_TIMEOUT_MS= # Per-LLM-request timeout (default 90000)
RSS_TIMEOUT_MS= # Per-feed fetch timeout (default 15000)
Client (build-time only):
VITE_API_URL — backend URL for a split deployment (e.g. https://api.example.com/api). Omit only if the frontend host proxies /api to the backend.VITE_API_TOKEN — must match the server's API_TOKEN when auth is enabled.
Securing a deployed backend
Running locally needs nothing: with API_TOKEN empty the API is open, and CORS
allows localhost only. As soon as the backend is reachable from anywhere
else, set both variables. Without a token, anyone who can reach the host can
delete categories, rewrite the system prompts that produce every summary, and
spend your LLM credits.
# server/.env
API_TOKEN=<a long random string>
ALLOWED_ORIGINS=https://your-frontend.example
# client build environment
VITE_API_TOKEN=<the same string>
GET /api/health stays unauthenticated so container health checks work. The API
also rate-limits: 300 requests/min per IP overall, 20/min on the LLM-backed
routes.
Prompt edits are preserved
Prompts you edit in the /prompts UI survive restarts. On boot the server
re-seeds the prompts defined in server/db.js, but only for rows still
byte-identical to what it last wrote (tracked via prompts.source_hash).
Anything you have changed is left alone and listed in the startup log. To
deliberately adopt the code versions, run once with PROMPT_SEED_FORCE=1.
Running
# Terminal 1 — Backend (port 3001)
cd server
node index.js
# Terminal 2 — Frontend (port 5173)
cd client
npm run dev
Open http://localhost:5173 in your browser.
Development Commands
# Type-check
cd client && npx tsc --noEmit
# Lint
cd client && npm run lint
# Production build
cd client && npm run build # outputs to client/dist/
Usage
- Add categories using the
+ button in the navigation bar (e.g., "World News", "Technology")
- Add RSS feeds by right-clicking a category and selecting "Manage Sources"
- Click Refresh to fetch and summarize the latest articles
- Browse history using the Archive panel in the left sidebar
- Chat about any summary using the chat panel below the article
- Send to Telegram using the button next to the Refresh button
- Browse jobs from the Jobs page (aggregated from 12 sources, configurable via
JOB_PROFILE)
- Discover new releases in the Movies & TV page
- Train your mind with the MindGames cognitive resilience dashboard
Project Structure
news-reader/
├── client/ # React frontend
│ ├── src/
│ │ ├── components/ # UI components
│ │ │ ├── ui/ # ShadCN UI primitives
│ │ │ └── *.tsx # Page-level components
│ │ ├── features/
│ │ │ └── mindgames/ # Cognitive Resilience Dashboard (9 sub-modules)
│ │ │ ├── common/ # Shared: FeaturePanelHeader, TabHeader
│ │ │ ├── dashboard/ # CognitiveDashboard, CognitiveTabNav
│ │ │ ├── overview/ # OverviewTab (stats, quick actions)
│ │ │ ├── training/ # InoculationPanel, PatternTests
│ │ │ ├── analysis/ # ForensicPanel, StudyStressTester, CompareCoverage, NewsSpectrum
│ │ │ ├── reflection/ # ScientistPanel, BridgePanel, JournalTrends, InformationDiet, StressDiagnostic
│ │ │ ├── reference/ # PromptLibrary, NarrativeMapPanel, DisinfoMap
│ │ │ ├── quiz/ # DailyQuiz
│ │ │ └── bias-radar/ # BiasRadarPanel, TechniquePicker, TechniqueCard
│ │ ├── hooks/ # useApi, useTheme, useWidgets, useMediaQuery, usePullToRefresh
│ │ ├── types/ # TypeScript interfaces
│ │ ├── utils/ # Date formatting utilities
│ │ ├── config.ts # API_BASE configuration
│ │ └── index.css # Theme definitions (4 themes via CSS custom properties)
│ ├── vite.config.ts
│ └── components.json # ShadCN UI config
├── server/
│ ├── index.js # Express entry point
│ ├── db.js # SQLite setup (WAL mode, auto-creates tables)
│ ├── routes/ # API route handlers
│ │ ├── categories.js # CRUD for feed categories
│ │ ├── feeds.js # Feed management
│ │ ├── summaries.js # Summary generation + history
│ │ ├── chat.js # AI chat about summaries
│ │ ├── briefing.js # Morning briefing generation
│ │ ├── jobs.js # Job board (12 sources + AI filter)
│ │ ├── stats.js # LLM usage + trending
│ │ ├── widgets.js # Weather, crypto, rates, HN, releases
│ │ ├── homepage.js # Newspaper grid homepage
│ │ ├── settings.js # User preferences (theme, etc.)
│ │ ├── telegram.js # Send-to-Telegram
│ │ ├── discovery.js # RSS feed auto-discovery
│ │ ├── narrative.js # Narrative map
│ │ ├── prompts.js # Prompt library
│ │ ├── disinfo.js # Disinfo map
│ │ ├── cognitive.js # Stress diagnostic
│ │ ├── forensics.js # Fallacy detection
│ │ ├── inoculation.js # Inoculation game
│ │ ├── scientist.js # ADEPT debate + journal
│ │ ├── bridge.js # SOS audit
│ │ ├── compare.js # Coverage comparison
│ │ ├── spectrum.js # News spectrum
│ │ └── bias-radar/ # Bias radar (decode, related, timeline, daily-quiz, steelman, missing-story)
│ ├── lib/ # Shared utilities
│ │ ├── llm.js # LLM provider fallback (Groq → OpenRouter)
│ │ ├── rss.js # RSS parsing
│ │ ├── telegram.js # Telegram bot integration
│ │ └── fetchWithTimeout.js # HTTP with timeout
│ ├── jobs/ # Job aggregator
│ │ ├── sources.js # 11 aggregator fetchers + ALL_SOURCES registry
│ │ ├── sources-ats.js # Direct Greenhouse/Lever/Ashby/Workable boards
│ │ ├── profile.js # JOB_PROFILE — configurable role/seniority/region
│ │ ├── ai-filter.js # AI relevance filtering (reads JOB_PROFILE)
│ │ └── common.js # Shared job utilities
│ └── middleware/ # Express middleware
├── AGENTS.md # Agent instructions & architecture reference
└── vercel.json # Vercel deployment config
Deployment
Both halves deploy independently. Whichever host you use, set API_TOKEN +
ALLOWED_ORIGINS on the server and VITE_API_TOKEN on the client — see
Securing a deployed backend.
Docker / CasaOS (what .github/workflows/deploy-*.yml uses)
Both client/Dockerfile and server/Dockerfile build standalone images; the
workflows SSH to the host and run ~/newsreader/deploy-frontend.sh /
deploy-backend.sh there.
- Backend — mount a volume for the DB and set
DB_PATH=/data/newsreader.db
(already the image default). Health check hits /api/health.
- Frontend —
client/nginx.conf proxies /api/ to http://newsreader-api:3001,
so put both containers on the same Docker network and the default
VITE_API_URL=/api works with no build arg. If your backend container has a
different name, change the set $api_upstream line in client/nginx.conf.
Frontend (Vercel)
- Build command:
cd client && npm install && npm run build
- Output directory:
client/dist
- Set
VITE_API_URL to your backend URL, including the /api suffix
vercel.json excludes api/ from the SPA rewrite. Without that exclusion the
rewrite matched /api/* too and every API call came back as 200 OK with an
HTML body — which surfaced as "Invalid response from server" rather than a
network error. The client now detects HTML responses on API requests and reports
the misconfiguration explicitly.
Backend (any Node host)
- Root directory:
server
- Start command:
node index.js
- Set
PORT (most hosts provide it)
- For database persistence: attach a volume (e.g.
/data), set DB_PATH=/data/newsreader.db
The process handles SIGTERM/SIGINT by checkpointing the SQLite WAL and
closing the DB, so restarts don't leave an unflushed WAL behind.
License
MIT
