Omni Viewer for Obsidian
Obsidian port of the vscode-omni-viewer extension. View (and in some cases edit) a wide range of file formats directly inside Obsidian.
Supported formats
| Category | Extensions |
|---|
| Archive | zip, rar, 7z, dmg, jar, apk, tar, tgz, gz, tbz2, bz2, txz, xz |
| Audio | mp3, wav, pcm, aiff, aif, aifc, amr, awb, ogg, flac, ac3, aac, m4a |
| Video | mp4, mts, m2ts, avi, mov, wmv, flv, webm, mkv |
| Image | jpg, jpeg, png, gif, bmp, webp, svg |
| Tabular | csv, tsv, xlsx, xls, parquet, arrow/feather (Arrow IPC / Feather V2), jsonl/ndjson/jsonlines |
| Font | ttf, otf, woff |
| Automotive / measurement | dbc, arxml, a2l, asc, blf, mf4, avro, bag (ROS), db3 (SQLite), reqif, pcap, pcapng, stp/step, h5/hdf5/he5, mat |
| Web / network | html/htm, har (HTTP Archive) |
| Jupyter Notebook | ipynb (nbformat v4) |
| eBook | epub |
| XML | xml, xsd, xsl/xslt, wsdl, xlf/xliff, rss, atom, xaml, kml, gpx, resx, csproj, vcxproj, nuspec |
| Documents | pdf (view + annotate/merge/save), docx/doc, ppt/pptx, hwp/hwpx, psd, md/markdown, tex/latex/ltx |
| Data / source | safetensors, gguf, onnx, tflite/lite, keras (and Keras HDF5 models saved as h5/hdf5), mlmodel/mlpackage (Core ML), pte (ExecuTorch), pt2 (PyTorch Export), OpenVINO IR (xml + bin), npy/npz (NumPy), json, yaml/yml, toml, proto, mmd/mermaid, puml/plantuml/iuml |
| GIS | shp (Shapefile), geojson |
Features carried over from the vscode extension
- Audio player with waveform/spectrogram (WASM-accelerated analysis for large files), regions, and region export.
- Image viewer/editor with filters and annotation tools; filtered images can be saved into the vault.
- PDF editor: text, stamps, signatures, page reorder, merge with another PDF, save / save-as.
- CSV editor: cell editing writes back to the file.
- HAR failure analysis: filter and sort requests by method, status, resource type, domain, URL, headers, and optional body content; inspect request/response details and page metadata alongside summary metrics and a timing waterfall.
- HTML preview: switch between a sandboxed preview and read-only source. Inline styles and embedded images are preserved; scripts, external resources, navigation, and active embedded content are blocked.
- Jupyter Notebook preview: Markdown and KaTeX math, highlighted code cells, execution counts, saved text/HTML/image/JSON outputs and error tracebacks, attachments and relative vault images, cell search, and source/output visibility controls. Read-only: code and interactive widgets are never executed.
- LaTeX structure and math preview: outline, source editing, tables, theorem blocks, KaTeX formulas, and vault-backed
\\input/\\include resolution. It is a partial preview, not a TeX compiler or typeset PDF.
- ML model inspection: ONNX, TFLite/LiteRT, and Core ML graph topology with searchable operator/tensor/IO tables and node inspection; ExecuTorch program instruction graphs with delegate, value, and segment inspection; PyTorch Export (
.pt2) graphs with signatures, weights, constants, and module stacks; OpenVINO IR topologies with constant previews from the neighbouring .bin; Keras layer tables, per-layer configuration, and parameter counts; GGUF metadata, tensor index, and quantization summary read without touching the tensor payload.
- EPUB reader: chapter navigation from the book's table of contents, text scaling, footnotes and in-book links, and a book-information panel. Chapters are sanitized before display; scripts in a book never run.
- XML inspection: element tree with attributes and namespaces, the source with the selected element highlighted, XPath copy, and search over names, values, and comments.
- Font inspection: local TTF/OpenType/WOFF specimens, editable sample text, size and variable font axis controls, Unicode character search, paginated glyph previews, and metadata/license details. Fonts are loaded only for the active viewer and released when it closes; WOFF2 and font collections are not supported.
How viewers are activated
Obsidian core already handles some extensions (md, pdf, images, audio, video, …). For those, this plugin cannot take over the default view; use the file context menu → Open with … Viewer instead. All other extensions (csv, ipynb, zip, parquet, dbc, hwp, xlsx, …) open with Omni Viewer by default.
Commands (Cmd/Ctrl-P):
Omni Viewer: Refresh viewerOmni Viewer: Share current fileOmni Viewer: Open shared link
Installation (manual / development)
npm install
npm test
npm run build
Then copy the following into <vault>/.obsidian/plugins/omni-viewer/:
main.jsmanifest.jsonstyles.css
The viewer templates, bundled JavaScript, and WASM assets are embedded into main.js during npm run build, so no extra asset folders are required.
Enable Omni Viewer in Settings → Community plugins. The same bundle runs on desktop, Android, and iOS.
Mobile support
On Obsidian mobile, Omni Viewer uses vault APIs instead of local filesystem paths. Mobile currently supports:
- ZIP/JAR/APK, audio/image/browser-native video, CSV, PDF, HAR, Safetensors, ONNX, TFLite/LiteRT, Keras, Core ML, ExecuTorch, PyTorch Export, OpenVINO IR, NumPy, JSON/JSONL, YAML, TOML, DBC, XML
- HTML, Jupyter Notebook, EPUB, Markdown, LaTeX, Mermaid, PlantUML, Protocol Buffer schemas
- XLS/XLSX, DOCX, HWP/HWPX, PPT/PPTX
Keras models open as .keras archives on mobile. Legacy Keras HDF5 models (.h5) rely on content-signature rerouting, which is a desktop-only path. Keras models are read into memory in full, so the 512 MB limit applies on both platforms; a legacy Keras HDF5 model above it stays with the HDF5 viewer, which reads metadata through the filesystem and opens at any size.
A .mlpackage is a directory bundle, so Obsidian only opens one that has been archived into a single file; a bare .mlmodel opens directly. Core ML models are read into memory in full on both platforms, so the 512 MB limit applies to each encoding.
ExecuTorch programs (.pte) are read into memory in full on both platforms, so the 512 MB limit applies; segment payloads (constants, delegate blobs) are located but never decoded. Separate .ptd data files are not opened.
PyTorch Export packages (.pt2) and NumPy files (.npy, .npz) are read into memory in full on both platforms, so the 512 MB limit applies. On desktop, a renamed .npy, a ZIP holding only .npy members, and a ZIP carrying a torch.export program are rerouted to these viewers by their contents.
XML documents open in the XML viewer, limited to 64 MB on both platforms. .arxml and .reqif keep the automotive viewers. XML saved under another viewer's extension (an Excel 2003 XML .xls, say) keeps that viewer.
HTML documents (.html, .htm) open on desktop and mobile, with a 64 MB vault-read limit. Core displays a bounded source preview above 8 MB or when markup is too dense to preview. UTF BOMs and early HTML charset declarations are supported. On desktop, recognizable HTML under an unknown extension also opens in the HTML viewer; extensions claimed by other viewers keep their existing routing.
OpenVINO IR shares the .xml extension, so an IR is recognized by its <net> root: desktop reroutes it to the OpenVINO viewer, and on mobile the XML viewer hands it over (an .xml up to 512 MB is read there to find out). A same-named .bin in the same folder is read for constant previews and range checks when it is 512 MB or smaller; the topology renders without it.
EPUB books are read into memory in full on both platforms and are limited to 256 MB; chapters and images are inflated on demand. On desktop, an EPUB under another non-archive extension is rerouted by its OCF mimetype member; one renamed to .zip stays in the archive viewer.
HAR files are read into memory in full on both platforms and are limited to 256 MB. For large captures, the viewer retains up to 20,000 requests and bounded request/response body previews; the core viewer reports when content was truncated.
Jupyter Notebooks (.ipynb) use nbformat v4 and are limited to 64 MB on both platforms. The core bounds previews to 2,000 cells, 10,000 saved outputs, 4 MB per preview, and 16 MB total preview content, with warnings when content is truncated. HTML/SVG outputs are sanitized, remote images are blocked, and relative images resolve inside the notebook's vault folder. On desktop, a v4 notebook saved as JSON also opens in the Notebook viewer when its nbformat and cells keys are present in the 64 KB detection window.
GGUF is desktop-first: models are inspected through filesystem range reads, so on mobile only files below 512 MB open (they have to be read into memory in full).
Files created by Save As, PDF merge selection, and archive extraction stay inside the current vault. Desktop-only native helpers are intentionally unavailable on mobile: RAR/7z/DMG/system-tar extraction, ffmpeg transcoding, LibreOffice PDF fallback, and legacy DOC rendering. Browser codec support can differ between Android and iOS.
Architecture notes
Each viewer is an Obsidian FileView. Core-backed viewers, including Jupyter Notebook, mount omni-viewer-core directly into the view and receive vault, navigation, and other host services from a small adapter. Notebook and KaTeX styles are scoped and bundled into styles.css. Legacy viewers host the vscode extension's self-contained HTML templates in an iframe; a bridge emulates acquireVsCodeApi() and handles messages with Obsidian/vault APIs. Both paths map theme variables to the active Obsidian theme.
License
MIT
