Readme

feedctl is designed to work out of the box. On first launch, you'll see a terminal-style boot sequence as the AI engine initializes. Once that completes, you're ready to go.

1. Pick your starter feeds. feedctl ships with a curated starter pack spanning technology, science, culture, and more. Or bring your own feeds from a different RSS app using OPML import.
2. Browse your queue. Articles are sorted by affinity score — the AI's calculation of how much you'll like each one. Higher scores appear first.
3. Vote on articles. Tap the upvote or downvote buttons as you read. This is how feedctl learns. The more you vote, the sharper the rankings become. You can vote in the detailed article view on the command bar, or in the list view by swiping left on an article.
4. Discover new feeds. Once you have a few votes, visit the Discover tab. feedctl recommends feeds by matching their content against your personal taste profile — entirely offline.

Boot sequence · first launch

Every article in your queue has an affinity score — a number that represents how closely it matches your demonstrated interests. This isn't keyword matching; feedctl uses semantic vector embeddings to understand the meaning of the text.

Here's what happens under the hood:

1. Each article's title and description are converted into a 384-dimensional vector using a local Core ML model This captures the semantic meaning of the text — not just keywords.
2. Your votes are also embedded. When you upvote "home automation with HomeKit" and "smart thermostat energy savings," feedctl recognizes the semantic overlap even though the wording is different.
3. For each new article, feedctl computes its similarity to your most relevant votes. High similarity to upvotes = high score. High similarity to downvotes = penalty.
4. Votes decay over time. Upvotes and downvotes are weighted by recency as your tastes naturally evolve.

Article queue · affinity scores visible

feedctl learns exclusively from your explicit votes. It never silently monitors your scroll position, dwell time, or tap patterns. If you didn't vote on it, it didn't happen.

• Upvote articles you find valuable, interesting, or want more of. Your upvotes pull similar future content toward the top of your queue.
• Downvote articles you find irrelevant, clickbait, or uninteresting. Downvotes push semantically similar content down your rankings.
• You don't need to read every article. Scanning headlines and voting on the ones that clearly hit or miss is enough to train the model effectively.

Article detail · vote buttons

Tapping an article in your queue opens the detail view. It's structured top to bottom so you can skim, commit, or dig in without hunting for controls.

• Header. Title, source feed, and publish date. Tap the title to open the original in your browser.
• TLDR button (iOS 26+, longer articles). A one-tap 2–4 bullet summary generated on-device by Apple Intelligence. Appears above the body when the article has enough content to summarize.
• Article body. If the feed only gives you a teaser, feedctl silently fetches the original page and extracts clean reader-view text — you'll see a brief Loading full article… indicator on first open, then the full prose. The expanded text is also re-embedded and fed back into the scoring engine, so the score sharpens as you actually read.
• Score Breakdown. Scroll past the body to see exactly why this article scored the way it did. On iOS 26+ you get a plain-English explanation ("matches your recent interest in distributed systems, similar to your downvote on cloud outages…") above the raw numbers: Attraction, Repulsion, an optional short-text penalty, and the Final score.
• Command Bar. A persistent footer with four actions: ▲ UPVOTE, ▼ DOWNVOTE, >> SAVE, and > OPEN. Voting trains the model (see §3); SAVE pins the article; OPEN hands off to Safari.

feedctl supports standard RSS and Atom feeds, plus Reddit subreddits as a first-class source. You can add feeds in two ways:

• Settings → Feeds — paste any RSS/Atom URL and hit Add. The feed fetches immediately and a toast confirms the result.
• Discover tab — browse feedctl's built-in catalog of curated feeds, matched and ranked against your taste profile. Tap to subscribe instantly.

For Reddit, use the standard subreddit RSS format:

Discover tab · personalized recommendations

feedctl's core scoring engine works on any supported device — the CoreML model runs via Core ML on all iPhones with an A-series chip or later. However, some advanced features use Apple's on-device large language model and require iOS 26+ on an Apple Intelligence-compatible device. You can check if Apple Intelligence is enabled in the left-drawer (INTEL: AVAILABLE / NOT AVAILABLE)

feedctl ships with four handcrafted themes based on the Catppuccin color system, each designed for extended reading sessions:

• Mocha — deep dark theme. The default. Low-strain, high-contrast text on dark surfaces.
• Mocha High Contrast — same palette, boosted contrast ratios for accessibility compliance.
• Latte — warm light theme with solarized parchment tones. Great for daytime reading.
• Latte High Contrast — light theme with sharper borders and bolder text weights.

Mocha · dark

Latte · light

• Vote broadly at first. The model converges faster with diverse signals. Don't just upvote your favorite niche — downvote the noise too.
• Steer with cluster chips. feedctl runs similarity clustering across your queue embeddings and surfaces the results as a horizontal chip bar above the article list. Tap a chip to filter the queue to a single semantic group; tap again to clear. On iOS 26+, each cluster gets a short natural-language label from Apple Intelligence (e.g. "AI tooling", "macro finance"); on older devices you'll see a top-term fingerprint instead.
• Push notifications track your baseline, not a fixed cutoff. When a background refresh brings in new articles, feedctl computes the mean affinity score of your current above-floor queue and only notifies you about new articles that beat that mean by a configurable multiplier. As your taste sharpens and average scores rise, the bar for a push rises with it — so you never get a flood. Tune the multiplier Advanced Settings.
• Full-text extraction is automatic. When a feed gives you a truncated teaser, feedctl fetches the original page the moment you open the article, extracts clean reader-view text, and re-embeds it so the score can sharpen against the actual prose — not just the headline.
• Your data stays local — always. There is no sync, no cloud, no telemetry. If you delete the app, everything goes with it.