RSSMonster – An agentic RSS reader built on local embeddings and small models RSSMonster, a self-hosted intelligent RSS reader developed by Piethein Strengholt, adds a semantic and ranking layer to traditional feed reading, grouping articles by event and personal interest while evaluating quality, freshness, originality, and source trust. The tool, which runs on local embeddings and small language models, offers declarative Smart Folders and explainable ranking decisions, with support for SQLite or MySQL for deployment. Copyright c 2026 Piethein Strengholt, piethein@strengholt-online.nl mailto:piethein@strengholt-online.nl RSSMonster is a self-hosted, intelligent RSS reader designed to help you cut through information overload and focus on what actually matters. Learn more about RSSMonster in the complete documentation. https://pietheinstrengholt.github.io/rssmonster/ Traditional RSS readers are primarily organized around feeds, folders, and chronological article streams. RSSMonster adds an intelligent semantic and ranking layer on top: it groups articles covering the same event and your personal interests, evaluates signals such as quality, freshness, originality, and source trust, explains why stories rank highly, and lets you create declarative Smart Folders for the views that matter to you. At its core, RSSMonster treats your feeds as a stream of signals rather than a pile of unread items. New articles are enriched with quality, freshness, originality, trust, attention, and semantic relationship metadata. That extra context lets the application answer better questions: is this worth reading now? , is this just syndicated copy? , which sources are covering the same event? , and which broader storyline does this belong to? A conventional reader effectively sees: Article Article Article Article Article Article RSSMonster can increasingly interpret that as: Topic │ Nintendo / Zelda │ ┌────────┴─────────┐ │ │ Event Related │ content ┌───┼───┐ A B C │ duplicates RSSMonster combines advanced search expressions, semantic clustering, quality analysis, and personal-interest-based rankings into a system where views are declarative, not hard-coded . Instead of fixed tabs and opaque algorithms, you define what matters using composable queries that power dynamic Smart Folders such as: Top Stories Today — importance-ranked, deduplicated coverage Worth Your Time — high-quality, original long-form content Quick Scan — summary-first daily overview Low Noise Mode — maximum signal, minimal volume Ranking decisions are explainable and views are customizable. The result is a reader that can behave like a quick daily briefing, a research inbox, a low-noise monitoring tool, or a classic feed reader depending on the view you choose. Semantic event discovery : RSSMonster groups reporting about the same real-world story into one expandable event, so several headlines from different sources become one event with multiple articles. Importance- and quality-aware ranking : Freshness, personal interest, article quality, breadth of coverage, source diversity, corroboration, and source trust help surface worthwhile stories without hiding the underlying signals. Declarative Smart Folders : Composable search expressions turn your own definition of “important” into reusable, dynamic reading views. Simple self-hosting : Run RSSMonster with SQLite and no separate database service, or use MySQL for larger and higher-concurrency deployments. Self-hosted, local, and transparent : Your feeds and reading data stay under your control, ranking dimensions remain inspectable instead of disappearing inside an opaque recommendation system, and pluggable small language models let the feed-processing pipeline run locally. Choose the reading experience that fits the moment, follow stories instead of duplicate headlines, and keep the same focused workflow across devices. Click any screenshot to view it at full resolution. Events and TopicsGroup related reporting into current stories and connect them to longer-running themes. | Interest IslandsSee the subjects your reading, favorites, and clicks keep reinforcing. | LandscapeA full dark-mode reading workspace on wider mobile and tablet screens. | PortraitA focused, touch-friendly article stream that travels with you. | The default Docker Compose deployment is designed for quickly seeing RSSMonster in live action. It uses SQLite, requires no separate database or model service, and starts the web application plus its dedicated crawl worker. For the comprehensive deployment—with MySQL and local inference using Qwen and ModernBERT—use MySQL Deployment mysql-deployment . git clone https://github.com/pietheinstrengholt/rssmonster.git cd rssmonster Create a .env file in the repository root: JWT SECRET=replace-with-a-long-random-secret FEVER CREDENTIAL SECRET=replace-with-a-long-random-secret Generate secure values with: openssl rand -hex 32 Run the command twice and use a different value for each secret. docker compose up -d The default docker-compose.yml is the quick live-action profile. It uses SQLite and stores the database in a persistent Docker volume. It disables inference-backed classifications, embeddings, the assistant, AI feed repair, and Smart Folder recommendations so it can start without downloading or running local models. On first startup RSSMonster automatically: - creates the SQLite database file; - initializes the database schema; - starts the application; and - starts a dedicated crawl worker that keeps due feeds updated. Open: http://localhost:3000 and create your first account. Check the deployment: docker compose ps The application validates database readiness, while the dedicated worker has its own crawl-health check. By default, three consecutive crawl failures or 15 minutes without a worker-state update mark the worker unhealthy. Follow the application and crawl-worker logs: docker compose logs -f rssmonster rssmonster-worker SQLite data is stored in the persistent Docker volume mounted inside the container at: /app/data The SQLite files can include: rssmonster.sqlite rssmonster.sqlite-wal rssmonster.sqlite-shm Do not remove the Docker volume unless you intentionally want to delete your RSSMonster database. To stop RSSMonster without deleting its data: docker compose down Avoid: docker compose down -v unless you deliberately want to remove the persistent database volume. The MySQL Compose deployment is the comprehensive RSSMonster profile. It is intended for installations that want higher write concurrency, multiple active users, and the local intelligent-content pipeline. It starts: - the RSSMonster web application, dedicated crawl worker, and rssmonster-ai-worker background-enrichment worker; - MySQL 8.4; - Qwen3 Embedding for 1024-dimensional semantic vectors; - Qwen3.5 for local classification text generation, Smart Folder recommendations, and feed rediscovery; and - ModernBERT for local article scoring. The comprehensive profile enables RSSMonster's AI-backed interface and processing features. No OpenAI API key is required for classification, embeddings, scoring, Smart Folder recommendations, or feed rediscovery. The optional natural-language assistant remains hidden unless INFERENCE ASSISTANT ENABLED=true is set after configuring ASSISTANT PROVIDER=openai and OPENAI API KEY , because its current inference adapter is OpenAI-only. Add the comprehensive deployment secrets and database passwords to the repository-root .env : JWT SECRET=replace-with-a-long-random-secret FEVER CREDENTIAL SECRET=replace-with-a-long-random-secret DB PASSWORD=replace-with-a-strong-database-password MYSQL ROOT PASSWORD=replace-with-a-different-strong-database-password Use the separate MySQL Compose configuration: docker compose -f docker-compose.mysql.yml up -d --build On the first startup, the inference container downloads Qwen and ModernBERT into the persistent inference-model-cache volume. This can take several minutes depending on the host and network connection. RSSMonster, its crawl worker, and its AI worker wait until MySQL is healthy and the inference models are loaded. Each worker reports its own health. Later starts reuse the downloaded models. Follow the complete deployment while it starts: docker compose -f docker-compose.mysql.yml logs -f inference rssmonster rssmonster-worker rssmonster-ai-worker Flexible reading modes : Use Reader Mode for summaries beside a details panel, List Mode for fast headline scanning, or Expanded Mode for distraction-free full articles. Keyboard shortcuts, drag-and-drop organization, dark mode, and mobile swipe gestures support efficient reading. Semantic event discovery : Group related reporting, compare sources, identify duplicate coverage, and connect events to broader topics and personal interest islands. Smart Folders : Build reusable views with queries such as @today unread:true sort:recommended , unread:true quality: 0.7 sort:quality , or event:true island:true eventCount: =3 sort:recommended . Advanced search : Combine article state, dates, tags, text, semantic filters, score thresholds, and sorting. See the search guide /pietheinstrengholt/rssmonster/blob/master/docs/search.md for the supported operators. Transparent ranking signals : Recommended ordering emphasizes personal interest, with freshness, Quality, corroboration, and rule tags as supporting signals. Top Stories separately ranks current multi-source event importance without personalization. Quality, uniqueness, attention, and feed trust remain inspectable signals where supported; attention sorting is retained only for legacy search expressions. PWA and mobile support : Install RSSMonster on supported devices for an app-like experience with offline support and responsive controls. OPML and generated RSS : Import or export subscriptions through OPML, and create filtered RSS feeds from stored articles through the /rss endpoint. Third-party client compatibility : Connect Fever clients such as Reeder or Google Reader clients including News+, FeedMe, Reeder, Vienna RSS, and ReadKit. Automated actions : Use regular-expression rules to delete, star, mark as read, flag as advertising, or mark matching articles as low quality. Multi-user support : Keep accounts, subscriptions, reading state, preferences, and assistant interactions user-scoped. Optional AI assistant : Enable natural-language search, summarization, classification, tagging, and feed interactions through the Model Context Protocol MCP . RSSMonster can notify a user when a completed crawl has persisted new articles, even when the installed web app is closed. Web Push is optional: RSSMonster continues to work normally when the VAPID variables are unset. VAPID identifies your RSSMonster server to browser push services. It uses one public/private key pair for the whole RSSMonster installation: VAPID PUBLIC KEY is sent to browsers when they create a push subscription. It is not secret. VAPID PRIVATE KEY signs outgoing push requests. Keep it secret and only provide it to the RSSMonster server. VAPID SUBJECT supplies operator contact information. Use a mailto: address or an HTTPS URL that belongs to the server operator. Each browser creates its own endpoint and encryption keys after the user selects Enable notifications . RSSMonster stores that subscription against the authenticated user. After a crawl, the server signs and encrypts a notification for each of that user's active browser subscriptions. The browser push service can route the encrypted message but does not receive the RSSMonster login token or VAPID private key. Keep the same VAPID key pair for the lifetime of an installation. Replacing it can invalidate existing browser subscriptions and require users to enable notifications again. Never commit the private key or paste it into client-side configuration. Install the server dependencies, then use the bundled web-push command: cd server npm install npx web-push generate-vapid-keys The command prints a public and private key. Copy them without adding quotes or whitespace. For a source installation, add them to server/.env : Optional Web Push notification configuration VAPID . VAPID PUBLIC KEY=replace-with-the-generated-public-key VAPID PRIVATE KEY=replace-with-the-generated-private-key VAPID SUBJECT=mailto:admin@example.com For Docker Compose, add the same values to the repository-root .env used by Compose: VAPID PUBLIC KEY=replace-with-the-generated-public-key VAPID PRIVATE KEY=replace-with-the-generated-private-key VAPID SUBJECT=https://rss.example.com Both included Compose configurations pass these optional values into the application container. Restart RSSMonster after changing them: docker compose up -d Restart a source installation after changing these values: cd server npm start - Serve RSSMonster through HTTPS in production. Browser service workers and Push subscriptions require a secure context; localhost is the development exception. - Install or open RSSMonster in a supported browser. On iOS and iPadOS, add RSSMonster to the Home Screen and launch that installed web app before enabling notifications. - Sign in, open the mobile Options sheet, and select Enable notifications . - Allow notifications in the browser or operating-system prompt. The control changes to Disable notifications after a subscription is active. It can also restore a missing subscription, remove the current browser subscription, explain unsupported or unconfigured states, and remove endpoints that a push service reports as expired. If RSSMonster says that Web Push is not configured, confirm that all three VAPID variables are present in the server process and restart it. If permission was denied, re-enable notifications through the browser or operating-system settings; a web application cannot reverse a denial itself. RSSMonster's newer architecture adds a semantic layer between feed crawling and the article list. Rather than storing articles as isolated feed entries, the system enriches them with vectors, scores, cluster membership, topic membership, and engagement signals. Those derived signals are then used by search expressions, Smart Folders, ranking, and the UI. The semantic pipeline works in stages: Article enrichment : crawled articles are normalized, summarized where applicable, scored for quality, and embedded into vectors that capture meaning beyond exact keyword overlap. Event clustering : each article is compared with recent candidate events using semantic similarity, headline overlap, named-entity overlap, and time proximity. Strong matches update an existing event; otherwise RSSMonster can create a new event cluster. Topic grouping : events are assigned to broader topics using ranked membership. An event can have a primary topic while still retaining secondary topic relationships, which keeps broad storylines stable without forcing every article into a single rigid category. Signal aggregation : event size, source diversity, topic density, freshness, quality, uniqueness, trust, and engagement are aggregated into ranking signals. This allows larger corroborated stories to surface without letting repetitive coverage drown out more original work. Declarative retrieval : Smart Folders and searches consume supported signals through visible query operators such as quality: 0.7 , freshness: =0.5 , event:true , island:true , hot:true , tag:security , and sort:recommended . This design keeps the intelligence of the reader inspectable. RSSMonster does not only decide what to show; it exposes the dimensions behind that decision so you can build views for different reading modes. A morning scan might prefer fresh event clusters with multiple sources, while deeper research might expand the full cluster, inspect related topic groups, and compare how different feeds covered the same story. Historical semantic rebuilding is available through npm run semantic:all . It rebuilds event, topic, and interest-island assignments for existing articles and is intended for explicit repair after large imports or algorithm changes. The visible sort order is Newest, Oldest, Top Stories, Recommended, Quality . Recommended : Emphasizes signed personal interest, then freshness and Quality, with small corroboration and rule-match contributions. It does not add a separate raw feed-trust preference boost. Top Stories : Ignores personalization and combines event coverage, cross-source diversity, corroboration, freshness, and Quality to surface broadly supported current stories. Article quality : Evaluates one article's writing, tone, and promotional content as an independent 0–1 signal. Quality ranking : Combines 70% article quality with 30% FeedTrust while keeping both concepts separate. Uniqueness : Describes how standalone an article is. Articles in larger event clusters receive a lower uniqueness signal, helping the interface identify redundant coverage without removing access to the underlying articles. Legacy sort:attention queries remain accepted for compatibility, but Most Engaged is no longer a visible sort option. Legacy sort:trust queries resolve to Quality. For the recommended Docker deployment: - Docker Engine or Docker Desktop - Docker Compose No separate MySQL installation is required when using the default SQLite deployment. For running RSSMonster directly from source: Node.js : Version 22.x or higher npm : Comes bundled with Node.js Git : For cloning the repository SQLite : Recommended for simple local and personal installations MySQL : Optional; recommended for higher-concurrency installations git clone https://github.com/pietheinstrengholt/rssmonster.git cd rssmonster Install server dependencies cd server npm install Install client dependencies cd ../client npm install Install inference dependencies cd ../inference npm install cd .. Copy the .env.example files to .env : cp server/.env.example server/.env cp client/.env.example client/.env cp inference/.env.example inference/.env RSSMonster sends all model requests to the standalone inference service. Configure the server connection in server/.env : INFERENCE URL=http://127.0.0.1:3001 INFERENCE TIMEOUT MS=30000 INFERENCE AI ENABLED=true INFERENCE ASSISTANT ENABLED=false SKIP ARTICLE CLASSIFICATION ANALYSIS=false SKIP ARTICLE EMBEDDINGS=false SKIP SEMANTIC LABELING=false Set INFERENCE AI ENABLED=false to prevent every server and worker inference request. This master switch overrides the feature-specific skip settings. Leave INFERENCE ASSISTANT ENABLED=false to hide chat while keeping the other intelligent features enabled. Set it to true on the server only after the assistant provider and credentials are configured in inference. Use a longer timeout such as 600000 when running Qwen on low-power hardware. The inference service selects providers independently for semantic embeddings, text generation, article scoring, and assistant responses. A complete OpenAI configuration in inference/.env is: OpenAI EMBEDDING PROVIDER=openai GENERATION PROVIDER=openai ARTICLE SCORING PROVIDER=openai ASSISTANT PROVIDER=openai ASSISTANT MODEL=gpt-4o-mini OPENAI API KEY=your-openai-api-key OPENAI EMBEDDING MODEL=text-embedding-3-small OPENAI EMBEDDING DIMENSIONS=1536 Alternatively, embeddings, article generation, and scoring can run locally while the assistant remains on OpenAI: Qwen and ModernBERT EMBEDDING PROVIDER=qwen GENERATION PROVIDER=qwen ARTICLE SCORING PROVIDER=modernbert EMBEDDING MODEL=onnx-community/Qwen3-Embedding-0.6B-ONNX EMBEDDING DIMENSIONS=1024 GENERATION MODEL=onnx-community/Qwen3.5-0.8B-ONNX GENERATION DTYPE=q4 ASSISTANT PROVIDER=openai ASSISTANT MODEL=gpt-4o-mini OPENAI API KEY=your-openai-api-key INFERENCE MODEL CACHE DIR=.cache/models Run inference with cd inference && npm run dev during development. Selected Qwen3 Embedding, Qwen3.5 generation, and ModernBERT models are downloaded and loaded during service startup, then reused from the model cache. The service logs when all configured models are ready and crawling can start. Development mode also logs content-safe activity for embeddings, summaries, tags, article scoring, assistant calls, Smart Folder recommendations, and feed rediscovery. Assistant responses currently continue to use OpenAI. See Model Usage /pietheinstrengholt/rssmonster/blob/master/docs/model-usage.md and Inference administration /pietheinstrengholt/rssmonster/blob/master/docs/inference.md for production setup and model-specific guidance. For a simple local installation, configure server/.env with: NODE ENV=development DB DIALECT=sqlite DB STORAGE=./data/rssmonster.sqlite RSSMonster creates the SQLite parent data directory when required. SQLite installations use conservative crawl concurrency settings automatically to reduce write contention. To use MySQL instead, configure: NODE ENV=development DB DIALECT=mysql DB DATABASE=rssmonster DB USERNAME=rssmonster DB PASSWORD=your database password DB HOSTNAME=localhost DB PORT=3306 Configure client/.env : VITE APP HOSTNAME=http://localhost:3000 Create the database schema: cd server npm run db If you explicitly need the project seeders: ./node modules/.bin/sequelize db:seed:all This section applies only to MySQL installations. When processing or querying large numbers of articles, increasing MySQL sort memory can reduce sort-related bottlenecks. Add the following to your MySQL configuration, for example in my.cnf : mysqld sort buffer size = 4M Run a crawl manually with: cd server DISABLE LISTENER=true npm run crawl This runs a crawl of active feeds and prints the crawl and semantic-processing results to the console. Production installations can run the dedicated crawl worker using the process-management approach appropriate to the deployment environment. If you need to rebuild article clusters from scratch: cd server npm run semantic:all This command rebuilds historical event assignments, topics, interest islands, and interest scores for every user. Use: npm run semantic:all -- --userId=3 to limit the rebuild to one user. When to use this: - after bulk importing articles; - when cluster quality degrades over time; - after changing clustering algorithms or parameters; - to repair cluster assignment inconsistencies. This is an explicit historical rebuild workflow. Normal post-crawl semantic processing only considers newly created, unfiltered articles. Taxonomy-vector generation is not required for a normal SQLite installation or Docker Quick Start . If you explicitly need to generate or regenerate taxonomy vectors: cd server npm run taxonomy:vectors npm run seed:island-taxonomy npm run taxonomy:vectors uses the embedding model selected by the running inference service, so it works with either OpenAI or Qwen. Feed trust estimates how consistently valuable a subscribed source has been as a source of articles: cd server npm run feedtrust This command calculates trust scores from 0.0 to 1.0 for active feeds using: Article quality 50% : Average existing normalized article quality Engagement 20% : Usefulness observed through reads, favorites, and clicks Originality 15% : Canonical articles versus deterministically linked duplicates Negative-feedback quality 15% : Explicit negative feedback among exposed articles When to use this: - periodically to update feed rankings; - after significant changes in reading patterns; - to identify low-quality or noisy feeds. Each signal has its own evidence confidence and shrinks toward the neutral score of 0.75 when evidence is sparse. Recalculating unchanged data produces the same result. Read the conceptual FeedTrust model /pietheinstrengholt/rssmonster/blob/master/docs/feedtrust.md . RSSMonster can expose an AI-powered assistant for natural-language interactions with your RSS feeds. It is optional and complements the core semantic pipeline rather than replacing event discovery, ranking, topics, or Smart Folders. Example requests include: - "Show me technology articles from the last week" - "What are my favorite articles?" - "Find unread posts about JavaScript" To enable the AI assistant and other OpenAI-backed capabilities, configure: Server server/.env : INFERENCE AI ENABLED=true INFERENCE ASSISTANT ENABLED=true INFERENCE AGENT TIMEOUT MS=300000 Inference inference/.env : OPENAI API KEY=your-openai-api-key-here ASSISTANT PROVIDER=openai ASSISTANT MODEL=gpt-4o-mini The server keeps no OpenAI credential; all provider calls go through inference. After configuration, restart the client, server, and inference processes. The assistant provides: - natural-language search across articles; - time-based filtering; - article summarization; - classification and tagging; - favorite and trending article discovery; - Smart recommendations based on reading interests. RSSMonster automatically tracks article interactions and can use AI to classify content with three quality metrics: advertisementScore — promotional or advertising-content detection; sentimentScore — emotional-tone analysis; qualityScore — content depth and writing-quality analysis. These scores provide additional inspectable signals for filtering and ranking. Note: All interactions are user-scoped, ensuring privacy and data isolation in multi-user environments. Note for Developers: The MCP server is available at /mcp for programmatic integration. Authentication requires a valid JWT token passed through the Authorization: Bearer