Why this exists

Most people who read news online spend an hour a day swivel-chairing — Hacker News, then The Verge, then Lobsters, then five team-specific blogs, then a podcast app, then back to the top because "did I check that one yet?". The platforms call this engagement. It's just lost time.

Algorithmic feeds (Twitter/X, Reddit, TikTok) solve part of the swivel-chair problem by aggregating, but they swap one cost for another: they decide what you see, optimised for the platform's engagement metrics, not yours. Your data leaves your machine. Your attention is sold.

Tech Feed Reader is the third option: aggregate everything you chose to follow, in your order, on your machine. Use AI to do the boring work — summarising, ranking, triaging — but never to decide what's important. That stays with you.

The anti-swivel-chair argument

The default RSS-reader UX (Feedly, NewsBlur) shows you every unread item in chronological order. With 50 feeds, that's a 200-item list at 8am every day. You either:

• Mark it all as read (the firehose problem)
• Skim every title and click into 10–15% (the time problem)
• Just stop using it (most people)

This app's answer:

1. Personal ranking. The For You sort scores every article against your bookmarks, thumbs-ups, and active reading time. Sort by relevance, not chronology.
2. AI triage. Once a night Claude reads your unread queue plus a sample of your positive + negative corpus and emits three groups: must-read, optional, skip. With rationales.
3. Summaries inline. Every article gets an extractive summary at import time. One click adds a Claude summary cached on the row. Skim mode hides the noise.
4. Skip the swipe. Bulk-action checkboxes; mute rules for keywords / authors / feeds; archive what you'll never get to.

How it works

• Ingest: a 10-minute cron pulls every subscribed feed via Ruby's FeedParser; HN/Lobsters teasers fall back to a readability fetch.
• Storage: SQLite with FTS5 for full-text search, WAL mode for concurrent reads. One file at data/app.db. No external services required.
• Ranking: a simple cosine-overlap score against your positive/negative corpus, plus a recency factor and per-feed weight that adapts from your engagement.
• Summaries: extractive (TextRank) at import, optional Claude (Sonnet 4.6) on demand. Cached so re-visits don't re-spend tokens.
• Triage: Claude reads up to 30 unread + 20 corpus exemplars per topic. Three runs nightly (cross-topic, technology-only, sports-only). Browse history at /triage.
• Sports: ESPN's unofficial API supplies scores, standings, fixtures, tennis rankings. Calendar exports as .ics for Apple/Google Calendar subscription.
• Audio: a Hotwire Turbo + data-turbo-permanent mini-player at the bottom of every page survives navigation; podcast episodes share the same article queue with a 🎧 badge.

What's different

• Single-user-first. No signup, no accounts, no auth boundary. The whole app trusts whoever opens it. (Multi-user is on the roadmap, but the simplicity comes first.)
• No analytics, no tracking, no third-party JS. The only outbound calls are the feeds you subscribed to + Claude's API for the summaries you opt into.
• Your data on your disk. Bookmark a backup of data/app.db and you're portable. Even the rotating background images are yours to swap.
• AI as a tool, not the master. Claude summarises and triages. It doesn't decide what you read, what's "for you," or what gets boosted.

Tech stack

• Ruby 3.4.1 / Sinatra / ERB / RSpec
• SQLite (FTS5, WAL mode)
• Anthropic Claude Sonnet 4.6 for triage + summaries; Opus 4.7 for chat widget
• Hotwired Turbo for SPA-style navigation; Chart.js for the Activity chart
• Cron-driven background jobs via Make targets (make refresh-feeds, make triage, make digest, make sync-sports)

Get started

Already running? Head to today's queue or browse feeds at /feeds. Ops stats live at /admin/dashboard. New here? The README in the repository covers make install and the seed-feed flow.