简体中文 · Downloads · Get started · Commands · Interactive gallery · Skills
A reader that thinks. The free and intelligent app for smooth agentic reading and beyond.
This repository publishes the Windows app, macOS previews, the interactive gallery, and the Skills that ship with ThinkReader. The desktop app source is not included.
ThinkReader is an agentic app for paper research. Your agents read alongside you and leave comments on the original PDF, so you can follow an idea back to its evidence without giving up control of what the paper really says. The knowledge you build together can support summaries, reports, mind maps, slides, and reproduction projects—all grounded in the source.
One kind of agentic paper reading turns papers into fast-food summaries; another floods the page with untraceable AI-generated information.
ThinkReader is neither: it keeps the original PDF, your judgment, and the agent's work in one place. It lets you open the paper, read it quickly but deeply, and turn what you learn into high-quality outputs. Only making the paper your own understanding can truly move your research forward.
flowchart LR
A["1. Upload a PDF"] --> B["2. OCR, semantic indexing<br/>and layout inference"]
B --> C["3. Understand together<br/>You × paper canvas × agents"]
C --> D["4. Create high-quality outputs<br/>Summary · report · wiki · more"]
A paper odyssey: you upload a paper to a Library; MinerU builds semantic and layout resources, while the original PDF always remains the reading surface. Then you, the paper canvas, and agents build understanding around the evidence together. That understanding can become a report, wiki, slides, mind map, or reproduction starting point—and always remains traceable to the paper.
- Read in flow. Agent annotations stream onto the PDF canvas without interrupting your reading.
- Multi-agent collaboration. Give agents distinct roles, work together in a Channel, and carry the conversation across reading and writing.
- Intelligent layout. Read in one to three columns and arrange the page around your needs; agents understand the original layout, making paper-grounded conversation effortless.
- Read once, reuse everywhere. Carry everything you understand while reading into summaries, reports, slides, mind maps, and reproduction projects.
- Find what's next. Compile knowledge across papers into a research network and find the next question worth exploring.
ThinkReader 0.9.70 is available for Windows x64 and macOS Apple Silicon.
This release fixes completion reporting for annotations limited to a selected section, reopening a recreated default Library, and detected figure, table, equation and algorithm borders. Detected regions now share the area-annotation theme style.
- Windows installer
- Portable Windows app
- Windows ZIP
- macOS Apple Silicon DMG
- Release notes, checksums and validation limits
Windows binaries are unsigned. Help & updates checks for new releases and can download the installer. Versions through 0.8.6, or older installations that restart without changing version, need one manual upgrade using the installer above.
The Mac app requires macOS 14 or newer and an Apple Silicon chip. It is ad-hoc signed and not notarized, so macOS may block opening it. Updates require manual installation. Intel Macs and automatic Mac installation are not supported by this release.
No terminal setup is required. On first launch, choose on-device parsing (local runtime and models) or the MinerU V4 API (API Token, Standard parsing; no CLI required); the App guides you through the required steps.
- On first launch, ThinkReader creates and opens a
defaultLibrary. Use the Library menu to create or open another Library. - Put your paper PDF in that Library's
RAW/folder, or import it from the Library home. - MinerU automatically processes imported papers. Local MinerU downloads models before its first extraction, so this step can take a while.
- Open the paper and start with
@Agent /browse,/summary, or/annotation. - Check the annotations against the source and cite a passage in your next question.
| Command | Purpose |
|---|---|
/browse |
Organize the paper's main thread into reading cards |
/annotation, /deep-annotation |
Annotate the paper at different depths |
/review, /deep-review |
Browse, annotate (or deeply annotate), then save a concise evidence-linked summary |
/audit |
Reply with an independent perspective on annotations; check and refine outputs |
/summary |
Create a concise summary for side-by-side reading |
/report |
Develop a longer research report |
/wiki |
Connect findings across papers |
/slides |
Prepare a research presentation |
/mindmap |
Organize ideas on a canvas |
/codebase |
Create a reproduction starter project |
/health-check |
Check Library records and paper-index consistency |
Paper commands default to the focused paper when each task starts. Explicit paths or paper IDs override that default; the Agent resolves natural-language titles and asks when the target is ambiguous. A running task keeps its selected paper. Reading commands default to the whole paper. An explicit section restriction limits the task, while a focus changes emphasis. Generated findings and code still need your review.
Attention Is All You Need · Explainable El Niño · Research atlases
Interactive demos show the way of desktop App running; agent execution in the online gallery is simulated.
System Skills used by ThinkReader commands live in skills/. Copy a Skill into the App's personal Skills folder to inspect or adapt it.
| Skill | Used by |
|---|---|
annotation-skill |
/browse, /annotation, /deep-annotation |
audit-skill |
/audit |
summary-skill |
/summary, /review, /deep-review |
report-skill |
/report |
wiki-skill |
/wiki |
slides-skill |
/slides |
mindmap-skill |
/mindmap |
codebase-skill |
/codebase |
app-control-skill |
/health-check |
A Library is a folder you choose in the App. Its main areas are:
Your Library/
├── RAW/ Original imported PDFs
├── .thinkreader/ Library identity, paper records and annotation sidecars
├── workspace/<paper-id>/
│ ├── mineru/runs/ Immutable extraction runs
│ ├── intelligence/ Semantic indexes, source geometry and reusable figures
│ │ └── document.md Editable reading copy
│ └── notes.md Independent paper notes, created when requested
├── output/ summary/, report/, wiki/, codebase/, slides/, mindmap/
├── index.html Library home entry
└── agents.html Agents entry
Some folders and files appear only when the corresponding feature is used. Keep the whole Library when backing up: editable notes and research outputs are user data, including those alongside derived indexes.
Markdown notes have base editable features. Drag tabs to arrange up to three reading columns. Themes, accent colors, and editable Skills adapt the workspace to your practice.
App settings and Agent state are managed separately from individual Libraries. Back up both when moving devices. Local storage does not mean local-only processing: selected content may be sent to the Agent or extraction provider you configure.
MIT License applies to the Skills and gallery in this repository.