User guide

A reader that pulls in what you follow, writes a summary of each piece, and keeps what you have read in step across your own devices.

Two things shape everything below. Your library stays on your devices. And a summary is written by a language model, which costs either your battery or your money depending on where you point it — so most of the settings here are about deciding how much of that happens without being asked.

This guide covers the desktop app. Screenshots are the Linux build in its light theme, taken over a demo library — one of its sources is a Nostr account, which is the one kind that does not work yet.

1. Getting started
2. The window
3. Reading
4. Summaries
5. Finding things
6. Keeping things
7. Managing sources
8. Sync
9. Keyboard
10. Settings reference
11. Where things are kept

Getting started

Adding your first source

SummaReader is empty until you tell it what to follow. Press + beside Your shelf, or Add source on the empty Inbox, and paste an address.

You can paste any of these and the right thing happens:

Nostr is the one that is not ready: the event parser exists, the relay transport does not, so an npub… is refused rather than half-followed.

The dialog offers both: Add source subscribes to something that keeps delivering, Save the page keeps one page and nothing more.

Importing what you already follow

Coming from another reader, Settings → Your data → Import and export takes an OPML file. Everything in it arrives as a source.

What happens on the first check

A feed's first poll delivers whatever archive it currently advertises — fifteen videos, thirty articles. None of that is new to you, so none of it is summarized automatically. Only things published after you subscribed are. Anything older can still be summarized whenever you want it: open it and press Summarize, or use the bulk button under Summaries.

The window

The top bar

Check for new and Filter appear while the list is on screen, which includes reading an article beside it. The Filter button fills in and reads Filtered whenever something is narrowing the list, so a list showing a fraction of itself always says why.

The shelf

Down the left: Inbox, Saved, Search, Sources, then your Lists, then your feeds grouped by kind. Each feed shows unread / total; each list shows what it holds.

• Click a feed to see only that feed.
• ⋯ opens that feed's own panel — see Managing sources.
• The eye beside + hides every feed with nothing waiting in it.
• « collapses the shelf to a strip of icons. b does the same.

The filter panel

Press Filter. Everything that decides what the list shows is here, and it is closed by default so the first thing under the heading is the first article.

Sorting cycles: Newest first → Oldest first → Title A–Z → Title Z–A → Feed A–Z → Feed Z–A, and back.

The read filter, the summarized filter and the sort are remembered between launches. Picking a feed is not — that is navigation, and opening into a library narrowed to one feed you chose last week reads as an empty app.

The counts under "Inbox"

Showing 18 of 5,878 · 52 unread · 812 to summarize

• Showing appears only when a filter is narrowing things, and counts the rows actually on screen.
• to summarize is the same number as the Summarize button, and is what pressing it would do.

Configuration

The handful of settings worth changing without leaving what you are reading: Text size (S/M/L/XL), List density, Thumbnails (None / Kind / Picture), line spacing and column width for the article, Reader layout, and Group by source.

List density decides how much of each item the list draws:

Thumbnails: Picture shows each article's own image instead of a kind icon. It is the only one of the three that costs anything: pictures are downloaded a few at a time for the rows on screen, so a long backlog fills in gradually rather than all at once. Dense ignores the setting — a picture is three lines tall, which is not dense.

Reading

Click an article, or press ↵. Depending on the window and your Reader layout setting, it opens beside the list, above or below it, or takes the whole window.

The three tabs

• Summary — what the model wrote, in three foldable sections: In short (four to six sentences), Key points (one per topic, each with a title and a couple of sentences explaining it) and Glossary (the terms it thought worth defining). Points from a video carry timestamps that seek. Shift+1–3 fold them.
• Source — the article itself, or the transcript.
• Details — where it came from, when, which model summarized it and what that cost.

1–3 jump between them.

Summaries written before this app stopped asking for one also carry The full notes, a longer prose version. Nothing produces one any more: what it was for went into the four-to-six-sentence overview and into the explanations under each topic.

The toolbar

• Save — files it into the list the shelf shows filled in, which is Saved until you choose another. The ▾ beside it, a right-click or a press and hold picks a different list or makes one.
• Mark read — on articles from a feed. With the unread filter on, this moves you to the next one; the last one closes the reader.
• Delete — only on articles you have kept. It asks twice, and the second press says Delete everywhere? because it removes the article from every device you sync with and stops any feed bringing it back.

Moving between articles

Drag the article sideways, or use j / k. Dragging right at the first one goes back to the list. The gestures are configurable under Settings → Reading.

When there is no text

Not every article arrives with its words. The reader says which of these it is:

Only the last is worth retrying, and Fetch the text does that.

Summaries

Where summarizing happens

Settings → Summaries → Where summaries run.

• On this device — a model you download. Nothing leaves the machine. It is slower and costs battery.
• On a server — your own, or a cloud API. Faster, costs money, and is the only time an article's text goes anywhere.

Whichever you pick gets its own tab beside Summaries, and only that one: Model for this device, LLM server for a server.

Choosing a model

Settings → Model. Search below the list, open a result's Variants to see the files inside it — they differ in size, and size is what decides whether one fits — and download the one you want.

Downloaded models are listed at the top. Start loads one and makes it the model that answers — starting it is how it is chosen. Only one runs at a time, so starting another asks first: the model in memory has to go before the next arrives. The choice is remembered, and Load it when the app starts brings it back on its own. Delete takes the files and the row together.

Some models will not run at all here — a file can require graphics support this build has no way to ask for, and the app says so rather than retrying.

Automatic summaries

Summarizing runs at the end of a check for new articles. There is no separate schedule — if Check for new is set to Only when I ask, nothing is summarized unless you ask. Two things decide what a check summarizes: each feed has a Summarize switch, off by default; and Settings → Summaries → At the same time sets how many one check may do, and which end of the queue it works through.

An article's own text is fetched during the check, not later — up to twenty per feed per round, newest first. An article that arrives without its words has usually failed rather than been deferred, and the reader says which.

Newest first is right day to day. If more arrives per check than the number allows, the ones it skips fall further behind every check — Oldest first is how you get back to them.

Every summary is a request. On a paid model, the number per check is what each check costs while you are not watching — worth choosing rather than arriving at.

Summarizing by hand

• One article — open it and press Summarize.
• One feed — its ⋯ panel, Summarize this feed.
• Whatever is filtered — the Summarize button in the filter panel. It says how many it will do, and that number is what it does.

What language a summary is in

By default, the language the article is written in. Settings → Summaries → Summary language overrides that if you would rather always read summaries in one language.

If nothing gets summarized

The most common cause is no model loaded. The toast says so, and Settings → Development → Activity log records the reason for every attempt.

Finding things

Search in the shelf, or /. Search covers titles, article text and video transcripts. A transcript hit is a moment — it names the time and seeks there when you open it.

The text box in the filter panel is a different thing: it narrows the list in front of you rather than opening results of its own. It matches the title, the source's name and the summary — Search also reads article text and transcripts.

Every filter asks the whole library, not the part of it on screen. A word, a tag, Unread, Not summarized, a feed, a kind: each one is answered from storage, so the number above the list is the number of things there are. tag:linux finds everything carrying that tag however old it is; two tags narrow rather than widen, and a value with a space goes in quotes: tag:"machine learning".

Tags

An article's tags are the subjects its summary picked out, and they are yours to change: the chips under In short in the reader add and remove, and clicking one shows everything carrying it.

A source can be tagged too, in its ⋯ panel — and an article is filtered by its source's tags as well as its own, so tagging a feed organises the hundred articles already in it rather than only the ones that arrive next.

Every tag in use is listed under Tags in the shelf, with what it holds. Clicking one narrows the list; clicking it again clears it.

Keeping things

Save in the reader, or s from the list. Saved articles are exempt from anything that clears space.

Lists

Saved is one list, and you can make as many more as you like — by subject, by what you mean to do with them, however you file. They live in the shelf on the left, under Lists, with what each holds beside its name.

Save goes to the list you last chose. Picking a list from the menu files the article and makes that list the one plain saving uses from then on — the shelf shows which by filling in its bookmark. Until you choose otherwise that is Saved, which is what saving has always meant.

An article can be on several lists at once, and taking it off one leaves the others alone. Saved itself cannot be renamed or removed: plain saving has to have somewhere to go, and every device agrees on that one.

Removing a list leaves the articles on it, as long as a feed or another list still holds them — the same rule as removing a feed. Removing a source leaves saved articles alone, and writes down where they came from before the source goes.

Lists travel between your devices, with the articles on them.

Managing sources

One feed

⋯ on a shelf row opens that feed:

Summarize again is what to use after changing the summary language or the model — an article that already has a summary is never redone otherwise.

Many feeds

The Sources screen is the same controls over the whole list, with a filter and a tick box, so "every YouTube channel I imported last week" is one action rather than three hundred.

Sync

Settings → Sync. Your devices keep each other in step: what you have read, what you saved, your highlights, and the articles themselves.

It is end-to-end encrypted: the server stores ciphertext and cannot read your library. There is no hosted service to sign up for — the address you set is a server you run.

When it runs

Settings → Sync → Syncing on its own offers a clock — every 15 minutes through to once a day — and two moments instead:

• After a check — as soon as every source has been fetched, before anything is summarized. On a local model summarizing can take twenty minutes, and the articles are ready long before that.
• After summarizing — once the whole check is finished, so one sync carries the articles and their summaries together.

Whichever you pick, a metered connection or a battery interval that has not elapsed still holds it back.

Pairing a second device

1. Turn sync on and set an address on the first device.
2. Pair a device shows a code.
3. Scan or type it on the second.

The first sync to a fresh server asks before uploading — a server that has never seen your library is a decision, not a detail.

The recovery code

Made once, under Sync. It is the only way back into your library if you lose every paired device. Nobody can reissue it.

Article text

Read state, highlights and the articles themselves travel automatically. Their words and pictures are the bulk, and by default they stay behind a pointer: sync tells this device that a copy exists, and it is fetched when you open the article.

Keep everything on this device changes that — each article's words and pictures are downloaded as they arrive. It is a setting per device rather than per library: worth it on a machine you read on offline, a lot of data on a phone for articles you will not open. Turning it on starts fetching whatever is already outstanding, a few seconds at a time, so a large library catches up over several syncs rather than in one long wait.

Keyboard

? shows this list in the app.

Settings reference

Reading

Theme, interface language, dates and times, reader layout, gestures, when an opened article counts as read, and whether the status strip is on screen. When an opened item counts as read is worth knowing about: Never means only you mark things read, Immediately means opening counts, and After a few seconds means passing through does not.

Summaries

Where summaries run, which model, how many at a time, how many per check and from which end, the summary language, and what summarizing has cost so far grouped by model.

Sync

Sync on or off, the server address, when it runs on its own, whether this device keeps everything, paired devices, the recovery code, and the local API — a door for a browser extension or another program on your machine to put something in.

Your data

Import and export: OPML for sources, and your library in full.

Development

The activity log — everything that ran and everything that failed, which is the first place to look when something did not happen. Also Fill in missing details, which re-fetches metadata, and Erase everything.

About

Version, and where the library file lives.

Where things are kept

Your library is one SQLite file in the app's data directory. Settings → About names the path. Backing that file up backs up everything: articles, summaries, highlights and what you have read.