mu-zi-lee/Tangentry

If your saved videos, podcast episodes and RSS articles live in five separate apps with five separate bookmark lists, Tangentry proposes one local reading inbox instead. You paste a link or subscribe to a channel, the content lands in a SQLite database on your own machine, and you read, search, annotate and export from a single interface. The friction it targets is real: platform feeds are built for consumption, not for retrieving what you learned three months ago.

The problem

A user wants to keep the useful part of what they watch and read: the argument in minute 14 of a lecture, a paragraph in a long article, a frame worth remembering. Today that means a browser tab, a platform favourite list, and a separate notes app with a pasted timestamp and no link back. Tangentry moves the note into the same place as the source, so a timestamp note, a highlighted paragraph, a PDF anchor or a captured frame stays attached to the item and can be searched together with everything else. The result is a personal corpus you can query rather than a pile of bookmarks.

What's different

The mechanism is a local Go server plus a React web UI, shipped both as a desktop app and as a Docker image for a NAS or home server. Content is normalised across Bilibili, YouTube, Xiaohongshu, Douyin, Xiaoyuzhou podcasts, RSS/Atom/JSON Feed and local PDFs; the reading layer adds subtitle following, Readability-style article extraction, playback position memory and optional machine transcription. The note model is the genuinely distinctive part: notes attach to a time point, a text range, a PDF location or a screenshot, and export as a Markdown ZIP with Obsidian frontmatter.

The trade-off is that this is a young 0.0.3.x project wrapping sites that change their interfaces. Video parsing and downloading rely on external yt-dlp, so breakage is expected to be fixed by updating that tool rather than the app itself, and some Bilibili content needs a login. AI summaries and translation require a user-supplied OpenAI-compatible endpoint, and cloud transcription sends audio to the provider you choose — the local-first guarantee applies to storage, not to features you explicitly route outward.

Who should try it

Good fit: a heavy consumer of long-form video and podcasts who already takes notes and wants those notes in Markdown next to an Obsidian vault, and who is comfortable running a Docker container or a desktop build. Poor fit: someone who mainly reads social feeds passively, or who wants a hosted service with zero setup and no dependency on yt-dlp, FFmpeg or a locally running process. The Docker route is the alternative for people who would rather keep one always-on instance on a home server than install a desktop app on each machine.

Requirements and limitations

Building from source needs Go 1.27+ and Node.js 24+, though prebuilt desktop and Docker routes exist. Downloads and frame capture need yt-dlp and ffmpeg; local transcription needs Python 3, and its speed depends on CPU/GPU and model size. Desktop auto-update requires an Ed25519 signing key configured in the repository, and the app must live in a user-writable location — installs under /Applications or package-manager paths will not self-update. macOS builds are ad-hoc signed by default and un-notarised builds can be blocked by Gatekeeper. Migration means copying the whole data directory while the app is closed.

HotRepo's take

HotRepo has tracked the project since 2026-10-24; its only observation is #16 on the NEW-REPOSITORY board (新项目榜, not the overall board) on 2026-10-25 with 146 total stars. That is a snapshot of a seven-day-old repository, not a growth series, so persistence cannot yet be judged. Before adopting, check the project's configuration and data-directory notes to confirm where credentials and the SQLite file live on your intended install path.