Capture. Read. Organize. Recall.

The default: the extracted text in one measured column, with the navigation, advertising and related-article furniture already stripped on the way in. Bionic emphasis is optional and off until you ask for it, and then it is yours to tune — how strongly the opening of each word is weighted, how much of the word it covers, and whether longer words get more emphasis than short ones.

• 

  One column, and nothing competing with it.

• 

  Bionic emphasis, with the weighting under your control. Placeholder.

Pacing moves the focus through the text instead of leaving you to keep your own place on the page. You choose the unit — paragraph, sentence or word — and how the focus is drawn: a background, an underline or a box. The surrounding text can be dimmed, a reading guide line can follow along, and the speed can adapt to how dense a sentence is, pausing at punctuation and holding longer on a heading.

• 

  The focus moves; you keep reading.

Book mode paginates rather than scrolls, with a table of contents, chapter headers and progress within the chapter as well as the whole work. On a wide screen it lays out a two-page spread, and an optional minimap shows where you are in the document. Comics get the same page-turning surface with rules of their own: the cover stands alone so every later pair matches print, and left-to-right or right-to-left is detected from the file at import and remembered on the comic rather than the session.

• 

  An imported ebook, paginated with its chapters intact.

RSVP presents the text a chunk at a time at one fixed point, so your eyes never travel. You set how many words appear at once, whether the flash pauses at punctuation, how many faded context words sit on either side, and whether the current sentence is shown underneath. The optimal recognition point inside each word can be highlighted, so every chunk lands in the same place.

• 

  One fixed point, and the text coming to you.

The reader persists a precise position, a progress percentage and a last-opened time as you go, and reopening restores the position rather than the top of the document. The save is deliberately withheld until the stored checkpoint has been read back: a reader that has just opened is sitting at the start, and writing that down would erase the place it is about to restore.

Select a passage to highlight it in a colour, attach a note to that exact selection, or drop a quick bookmark. These are durable library data rather than reader session state: they survive reopening, they are searchable across the whole archive, they become nodes in the knowledge graph, and they are collected in a Highlights section so you can read back what you marked without reopening each item.

• 

  A passage marked, with a note on the selection itself.

• 

  Everything you marked, collected across the archive.

Search inside what you are reading, step between matches, and jump to the chapter or block holding one. It is deliberately separate from archive-wide search — and when an archive search is what opened the item, its anchor is carried into the reader, so you land on the passage that matched rather than at the top of a long document.

• 

  Matches counted, and a jump straight to the chapter. Placeholder.

Anything with real chapters keeps them: a working table of contents, chapter headers, progress inside the current chapter, and a resume that knows which chapter you were in. That covers EPUB, the Kindle formats, FictionBook and Royal Road serials. Comics have pages rather than chapters and use the comic surface instead.

English is spoken by a neural voice model that runs on your device. You install it once from settings, and after that nothing about what you are reading is sent anywhere to be spoken. Other languages use the voices your operating system already provides, picked per document rather than per app. The sentence being spoken is highlighted as you listen, so you can follow along or take over reading at any point: select a passage and choose Read from here to start where you want, and once it is reading, click any word to jump there. Speed runs from half to triple. It is free, needs no account, and runs in Normal mode — starting it from Book, Pacing or RSVP switches the reader there first.

• 

  The sentence being spoken, highlighted as you listen. Placeholder.

• 

  One download, then a voice and a speed that stay yours. Placeholder.

The library, your reading positions, your notes and the search index all live on the device, and none of them ask a server for permission to open. Anything already cached reads exactly the same with the network off. Turn sync on and it carries enough cached content that an item saved on another device opens on this one too.

The assistant answers from the item in front of you and nothing else. The retrieval boundary is explicit rather than implied — document chat can reach the open item and the passages attached to it, and archive-wide access is only available from the archive scope, which is labelled as such. It can read around your current position, so "explain this" means the part you are actually looking at. It runs on your own provider key, or on a local model.

• 

  The assistant, before a provider of your own is configured.

An answer carries citations back to the archive items behind it, with the excerpt and, where there is one, a search anchor. Following a citation opens the reader at that passage rather than at the item. That is what makes the assistant a route back into your own reading instead of a second thing to read.

• 

  An answer that hands you back the passage it used. Placeholder.

One library, with sections for what you have not started, what you are in the middle of and what you have finished, plus an inbox-style entry point for what has just arrived. Every item has a detail panel carrying its metadata, notes, collections, tags and the history of any rule that filed it. List, grid and masonry are the same archive seen three ways, not three places to keep things.

• 

  The archive as rows: status, source and date at a glance.

• 

  The same archive as cards, when you are browsing rather than working.

• 

  An item's detail panel: progress, length, source, collections.

Collections nest, so a project can hold its own sub-collections instead of flattening into a soup of tags, and items move by being dragged. An item can belong to more than one collection, and adding it to one never takes it out of the archive — nothing is filed away by being organized.

• 

  Collections inside collections, arranged by dragging.

Tags cut across collections rather than competing with them: a collection is where something lives, a tag is what it is about. The same tag works as a filter, as a Smart Rule condition and as a node in the knowledge graph, so the effort of tagging pays back in more than one place.

Narrow the archive by content type, reading status, date, whether an item carries notes, and more. Filters compose with the section you are in and with search, so "unread PDFs I annotated last month" is a few clicks rather than a query language you have to learn first.

• 

  Narrowing a large archive without leaving it. Placeholder.

Select many items and act on all of them at once — tag them, file them, mark them read — from the toolbar or a context menu. The same selection behaves identically in list, grid and masonry, so tidying up does not mean first finding the view that lets you.

• 

  Many items, one action. Placeholder.

The index is built on your device out of what you actually saved: titles, authors, tags, your notes, the text of your annotations, chapter titles and the cached full text of documents. Results carry a snippet and a search anchor, so opening one puts you at the matching passage rather than at the top of a fifty-page document. Nothing is uploaded to be indexed, and search works with the network off.

• 

  A match, and a way straight back to the passage that matched.

A rule is a tree, not a single line: nested AND/OR groups with negation, matching on source, author, title, tags, length, date, reading behaviour and knowledge-graph attributes. Actions are additive and idempotent — add to a collection, add a tag, mark as read — so a rule can never quietly undo something you did by hand. Rules can be tested against the real archive before you commit them, so you see what one would have done rather than finding out afterwards, and every item keeps an audit history of the rules that touched it.

• 

  A condition tree, tested against real items before it is saved.

You invite specific people by email address. There is no public link, and nothing is shared until the recipient accepts from their inbox. Each person gets a role and up to five separate permissions: edit notes, edit tags, edit annotations, add documents, remove documents. A read-only invitee genuinely cannot write, and that is enforced on the server as well as in the app, so it is not merely a hidden button.

• 

  Named people, named permissions, and no public link. Placeholder.

A collection can leave as a single .flowcollection file carrying its items, tags, annotations and reading positions, importable by anyone running FlowReader with no account on either side. HTML export is the other direction: readable anywhere, no FlowReader required. It is also the honest answer to "what if I stop paying" — your reading is a file you already have.

• 

  Everything leaves in one action, in a format that outlives the app.

Chat bounded to one collection and its children: retrieval cannot reach outside it, and answers cite the items they used. It is the difference between asking a model what it knows and asking your own reading what it says.

Set a target and watch it against what you actually read. Goals draw on the progress the reader is already recording, so there is nothing to log by hand and nothing to keep honest.

• 

  A target, measured against real reading. Placeholder.

Progress per item and per chapter, and reading activity over time. It is computed locally from your own reading and stays on the device unless you turn sync on.

• 

  Reading activity over time, computed on the device.

A Flow starts with a request, not a configuration wizard. It can work from watchlists, your archive or open-web discovery; collect new items, track a changing signal, write a cited briefing, or combine those outputs. Deterministic collection and compatible signals run without an AI key. Agentic research uses your configured provider and states whether it is running locally or on the FlowReader server.

Built on your device from the archive you already have. There is no upload step and no separate service. Every item, tag, collection, annotation, author and source becomes a node, and the relationships between them become edges — so it gets denser as you save, rather than as you maintain it. The baseline needs no AI provider at all: entity extraction and TF-IDF relatedness are local calculations. A rebuild is checkpointed, so interrupting one does not start it over.

• 

  The overview across a whole archive.

• 

  People, organisations and projects, pulled out of what you saved. Placeholder.

• 

  Why two items are related, shown beside the one you are reading. Placeholder.

• 

  Clusters the reading formed on its own. Placeholder.

• 

  The items sitting between two separate areas. Placeholder.

• 

  An answer assembled from the graph, citing the archive behind it. Placeholder.

Every item entering the archive goes through a heuristic extractor that pulls out what is being written about: people, organizations, projects, identifiers. It is deliberately conservative — it publishes evidence for explicit identifiers, distinctive names, and proper names that either repeat or are supported by the title, rather than guessing — and it is anchored, so a claim is tied to where in the text it came from. It needs no AI provider. Optional enrichment can classify what the heuristics left unlabelled, queued, resumable and capped per day, and anything you curate by hand stays authoritative over anything a model suggests.

Community detection over the graph surfaces the groups of items that hang together: the subject you have been circling for two months without ever making a collection for it. They are found rather than declared, which is the point — you do not have to have been tidy for them to appear, and a cluster that proves itself is a good starting point for a real collection.

Relatedness starts with TF-IDF similarity over the text you actually saved: a local calculation, no model and no network. Turning on embeddings adds a semantic signal that catches two items about the same thing in different words. Both feed the same edges, and a related panel sits inside the reader, so a connection surfaces while you are reading rather than only when you go looking for one.

Centrality analysis finds the items sitting between clusters: the paper that links your security reading to your distributed-systems reading. Those are usually the most valuable things in an archive and the hardest to find by browsing, because they belong neatly to neither side.

Items read in the same sitting become connected to each other. It captures something no amount of tagging does: that two pieces belonged to the same question on the same afternoon, whatever their subjects were.

With your own key configured, the graph stops being something you look at and becomes something the assistant can query: trace the path between two entities, summarize everything your archive says about one, find every item mentioning it, list the topic clusters or the bridge items, or ask what to read next based on structure rather than on what arrived most recently.

A skill is a named instruction set plus a bounded list of the tools it may use. Pick one in the composer, or let the assistant invoke it by name. The built-ins cover summarizing, explaining, fact-checking with citations, translating, quizzing yourself for retention, flashcards, critiquing an argument, comparing sources, deep research, graph exploration, reading recommendations and writing — and every scheduled Flow briefing is powered by one too. Editing a built-in forks your own copy rather than overwriting it, and you can write one from scratch.

• 

  A skill: instructions, plus the tools it is allowed to use. Placeholder.

Background analysts for work too large for a single answer: analysing a whole collection, comparing two items, exploring a topic, or working through a range of chapters. They run alongside the conversation rather than blocking it.

Connect a third-party Model Context Protocol server and its tools become available inside FlowReader’s own chat, named so you can always tell them from the built-ins. Connections are HTTPS only, the credential lives in the volatile store rather than in the settings table, and everything a server returns re-enters the prompt fenced as untrusted data. This is FlowReader reaching outward; letting an AI client reach in is the other direction, and it has a chapter of its own on the Pro page.

• 

  A third-party tool server, bridged into the assistant. Placeholder.

Point FlowReader at Ollama or LM Studio and the assistant runs on your own device: no key, no provider account, and no request leaving the device. It works against a local model the same way it works against a remote one — the difference is the speed and the quality of whatever you are running.

Anthropic, OpenAI, Google and Hugging Face routing, with the key kept on the device and never synchronized to another device or to a server, on any provider. There is no FlowReader-owned key anywhere in the product, which is why AI is labelled as needing your own credentials wherever it appears: the cost is yours, it is visible, and it stops when you stop.

• 

  Your provider, your key, or no provider at all. Placeholder.