English · 简体中文
A local-first ebook reader built around Focus Mode—one meaningful passage at a time.
Native Rust rendering, no WebView, and your library stays under your control.
Focus Mode • Features • Screenshots • Download • Privacy • Development • Project status • License
Torto is an open-source desktop reader designed around Focus Mode. Instead of treating a book as a stack of pages, Focus Mode turns its structure into meaningful reading units—paragraphs, nested lists, quotations, code blocks, tables, and images—and keeps the current unit in a stable reading position. Navigation, highlights, notes, footnotes, translation, and AI conversation all follow that context.
Classic single-column, two-column, and vertical-scroll layouts remain available when you want a conventional page or chapter view. The local bookshelf, full-book search, translation, optional AI assistant, PDF OCR, and direct WebDAV sync are built around the same native reading model.
Unlike browser-based readers, Torto parses, lays out, paginates, and renders book content through a native Rust pipeline built on egui, Parley, Vello, and wgpu.
Focus Mode is Torto's default and flagship reading experience—and the feature that most clearly sets it apart from conventional ebook readers.
- Read by meaning, not by page. One semantic unit is active at a time. Images are valid reading units, nested list descendants stay with their parent item, and large tables or media remain scrollable instead of being skipped.
- Browse lists with their context. A list includes its immediately preceding ordinary paragraph. A fitting group gets one green rounded activation; a taller group has a subtle gray background and green activation for its introduction or current top-level item. Oversized items scroll internally before navigation advances, and list footnotes share one numbering scope.
- Keep your place visually. Short units begin at a stable position inside a centered reading stage; taller content expands only when it needs more room. Moving between units avoids the constant vertical jump caused by paragraphs of different lengths.
- Navigate consistently. The mouse wheel, arrow keys, and the reader scrollbar move through reading units. When a table of contents, chat, footnote panel, or text editor owns input, scrolling stays inside that surface.
- Act on the current passage. Open the action toolbar with
Space, chat withTab, highlight with1, add a note with2, use Split by sentence with3, and toggle contextual footnotes withLeft Alt. Shortcuts are configurable. - Keep context attached. Focus chat sessions are scoped to their reading unit for the current book session, while highlights and notes persist at stable source locations. Footnotes are reduced to compact icons on the active unit and expanded only on demand.
- Bring scanned PDFs into the same workflow. Original-layout PDFs use Classic Mode. After PDF OCR creates a reflowable text layer, Focus Mode becomes available with the same unit navigation and reading tools.
| Feature | Description | Status |
|---|---|---|
| Focus Mode | Read one semantic unit at a stable position, navigate by unit, and keep actions, conversations, notes, and footnotes attached to the current context. | ✅ Core |
| Multi-format support | Read DRM-free EPUB, MOBI, AZW, AZW3/KF8, FB2, FBZ, CBZ, CHM, and PDF files. | ✅ |
| Native rendering | Parse, lay out, paginate, and render books with Rust instead of embedding a browser or WebView. | ✅ |
| Classic layouts | Switch between single-column, two-column, and section-based vertical scrolling layouts, with unified typography or the book's authored styles. | ✅ |
| Cover-first library | Import books in bulk, browse cover cards, search by title or author, resume recent reading first, and detect duplicates. | ✅ |
| Navigation and search | Use a hierarchical table of contents, chapter tracking, full-book search, keyboard navigation, mouse-wheel paging, and F11 fullscreen. |
✅ |
| Typography and themes | Configure default, CJK, code, and interface fonts; choose unified or book-authored typography; and follow the system, Light, or Dark theme. | ✅ |
| Selection and annotations | Select freely or by word, sentence, or paragraph; copy text, create highlights and notes, and return to durable source locations. | ✅ |
| Image preview | Press 0 on an active image in Focus Mode, or open it with the pointer. Images initially fit the window, including enlargement of small images; zoom with - / = or the wheel, pan when enlarged, and copy images to the clipboard. Recognized formulas preview their rendered LaTeX. |
✅ |
| Translation | Translate book content in replacement or bilingual mode and translate the table of contents. | ✅ |
| AI layout | Recover missing headings, attributed quotations and captions; turn inline citations into popup references; recognize text and image formulas and render them as mathematics. | ✅ Optional |
| References and links | Read footnotes and inline citations in a shared popup; open website icons in your browser in unified typesetting. | ✅ |
| AI reading assistant | Stream answers with source-backed citations and render Markdown, tables, math, SVG, and Mermaid, including enlarged visual previews. | ✅ Optional |
| PDF OCR and metadata recognition | Recognize title, author, table of contents, page roles, and body content; switch between the original PDF and a reflowable OCR layout. | ✅ Optional |
| WebDAV sync | Sync books, reading progress, highlights, and notes directly through your own WebDAV provider. | ✅ Optional |
| Windows updates | Check GitHub Releases automatically, verify the MSI with SHA-256, and install updates after confirmation. | ✅ Windows |
Select an AI layout model and reasoning effort in Settings → Typesetting → Layout. AI layout supplements the book parser: it can identify headings stored as ordinary paragraphs, quotations with explicit sources, missing attribution for existing quotations, and figure or table captions. Tables represented by images are supported too.
Inline citations appear as numbered markers, restarting within each paragraph. In Focus Mode, their popup lists footnotes first, then inline citations. Formula recognition handles both images and formulas written as text, and copying a recognized formula yields its LaTeX text. Unified typesetting also turns website addresses, including common bare domains, into clickable internet icons.
Recognition prioritizes the current reading area and processes the active subsection in stable batches. When translation is enabled, AI layout follows its reading range and prepares semantic markup before translation; related results are applied together to reduce layout jumps. Recognition can be imperfect, and rejected formula conversions retain the original content.
In chat, type / for reading skills or @ to attach references. Use the arrow keys to select a suggestion and Tab to insert it; when suggestions are open, Tab keeps the conversation open.
PDF contents, book metadata, and special pages are discovered by a shared tool-using agent. It chooses which page overviews, full pages, crops, and existing text to inspect, saves an incremental draft, and submits results after local checks. Printed labels and physical page numbers are stored separately; no single book-wide page offset is assumed. Interrupted drafts and rendered page images are cached locally for reuse. Only confirmed navigation targets are applied; partial completion is reported explicitly. This feature uses the configured recognition model and requires both vision and tool calling. Each run is bounded to 20 model turns and ten minutes, with at most five full pages per read and bounded context history.
Expert translation is an optional switch in Settings → Translation, off by default. It collects up to eight useful new terms alongside each translation, keeping the existing paragraph-keyed response and adding a compact g array of s (source) / t (translation) pairs. Terminology is saved locally per book and target language; subsequent requests receive only matching terms, within a bounded prompt budget. Extraction excludes ordinary phrases and widely established names such as Steve Jobs and McDonald's. Conflicting suggestions never overwrite established terms, and invalid glossary metadata does not discard valid translations. Enabling the switch affects future requests without automatically retranslating existing text; disabling it retains the glossary for later use.
The searchable provider selector includes OpenAI, Anthropic, Gemini, DeepSeek, OpenRouter, SiliconFlow, xAI, Groq, Mistral, Moonshot, MiniMax, Z.ai, Ollama, llama.cpp, and custom OpenAI-compatible endpoints. Requests use Rig adapters behind a shared transport layer. Existing provider IDs, credentials, model selections, and custom gateway URLs are preserved. Model lists use the provider's protocol; model IDs can also be entered manually.
For structured tasks, callers supply one JSON Schema. Automatic output selection tries native JSON Schema, then schema-based tool calling, JSON object mode with a schema prompt, and finally prompt-only JSON. Unsupported strategies are skipped or remembered temporarily for the matching endpoint, model, schema, and request capabilities. Authentication, rate-limit, and server errors do not trigger format fallback. Every result is parsed with the existing JSON repair fallback and validated locally before use. Output selection is automatic and requires no provider configuration; custom providers use prompt-only JSON by default. Structured tasks share a four-request budget across format fallback and existing correction attempts.
Import books, browse covers, search by title or author, and continue where you left off.
Focus Mode is the default experience described above. Torto also keeps a conventional two-column layout for readers who prefer a page-like view.
Download the latest build from GitHub Releases.
| Platform | Package | Requirements |
|---|---|---|
| Windows | Torto-*-x86_64.msi |
64-bit Windows 10 or 11 |
| macOS, Apple silicon | Torto-*-macos-arm64.dmg |
macOS 15 or later |
| macOS, Intel | Torto-*-macos-x86_64.dmg |
macOS 15 or later |
After installation, import one or more books from the bookshelf and open a book in Focus Mode. Use ↑ / ↓, the mouse wheel, or the reader scrollbar to move between reading units; press Space for passage actions and Tab for passage-scoped chat. Use Ctrl + F to search the current book and the reader menu to change layout, typography, theme, translation, AI, OCR, shortcuts, and sync settings.
- Imported books and reading data remain in Torto's local application-data directory.
- WebDAV passwords and AI API keys are stored in the operating system's secure credential store rather than ordinary configuration files.
- AI layout, translation, chat, and OCR use your configured providers. Once enabled, AI layout and translation can automatically send relevant book text or images as you read, including subsection context or upcoming content within their scheduling range.
- WebDAV traffic goes directly from Torto to the service you choose; there is no Torto-operated relay.
Torto uses Rust 1.97.1. The workspace package remains named rebook-desktop, while the shipped application is Torto (torto.exe on Windows).
cargo run --locked -p rebook-desktop
cargo fmt --all --check
cargo clippy --workspace --all-targets --all-features -- -D warnings
cargo test --workspaceThe core reading path is parser → Reading IR → layout → renderer. See the native rendering architecture decision, AI layout batching, AI layout and translation scheduling, WebDAV sync protocol, and known upstream issues for details.
Torto is under active development. DRM-protected books are not supported. Focus Mode is unavailable for an original-layout PDF until reflowable OCR content exists. The native renderer intentionally does not aim for full browser-level HTML/CSS compatibility, so complex fixed layouts, vertical writing, Ruby annotations, and some interactive book content may not render completely yet.
Please report reproducible problems in Issues. Include the book format, screenshots, and reproduction steps, but do not upload complete copyrighted books.
Torto is available under the GNU Affero General Public License v3.0 (AGPL-3.0-only).