Transcript Desk

A local research desk for the internet. Paste a YouTube, TikTok, or Instagram link, a tweet, or an article URL — or drop in video, audio, or a screenshot — and get clean text (timestamped transcript, tweet + OCR'd images, or extracted article), an AI summary with chapters, and a research pack (claims table, fact/opinion split, verification queries), organized in a searchable library that also lives as plain Markdown on your disk.

Everything runs on your own Mac. Transcription is local (Whisper on Apple Silicon), fetching is local (yt-dlp), storage is plain files. The AI layer runs through the Claude Code CLI on the Claude subscription you already have — no API keys anywhere.

Why this exists

Free transcript sites are a dime a dozen. What they don't do: keep a library, run privately, work on TikToks and reels, or take the transcript seriously as research material. Transcript Desk's reason to exist is the layer after transcription — persistent research packs with timestamped claims you can trace back to the second they were said, folders that mirror your projects, full-text search across everything you've ever transcribed, and a Markdown archive that outlives the app.

What it does

• Paste a link (YouTube, TikTok, Instagram, podcasts — anything yt-dlp speaks) or drop files anywhere on the window, several at once
• Captions-first: if the platform has captions, the transcript is instant; otherwise audio is fetched and transcribed locally with Whisper (large-v3-turbo via mlx-whisper — free, private, multilingual)
• AI summary with timestamped key takeaways, chapter titles on long content, translation (Romanian ⇄ English by default — one prompt to change), and an on-demand research pack: claims table, exact quotes, fact/opinion/speculation split, suggested verification queries
• Playlists: paste a playlist link, every video becomes its own note, processed one at a time
• Tweets: text, images (read via Apple's Vision OCR — local and free), and quoted tweets captured; video tweets get the full transcript pipeline; link-only tweets capture whatever they point at
• Articles: readability extraction to Markdown (Defuddle), with a headless-Chrome fallback for script-rendered pages
• Screenshots: drop any image — the text is read out of it
• Reading list: a separate section for books — add by title and the author, description, and tags fill themselves in; copy or export as Markdown
• Library: folders, full-text search with timestamped snippets, real thumbnails (uploaded videos get a frame grab, audio gets a waveform tile)
• Markdown mirror: every note is also a formatted .md in ~/Documents/Transcripts/ — YAML frontmatter, folders mirrored as subfolders, deletions archived to _history/, readable in Obsidian
• Exports: .md / .txt / .srt, copy buttons, and a one-tap "Copy Claude prompt" for taking a transcript into a chat for deeper work
• Phone: open the same app from your phone's browser and Add to Home Screen — the phone is a live window into your Mac's library, so there's nothing to sync. Pair with Tailscale and it works away from home too, without exposing anything to the internet.
• Light and dark theme (follows the system, manual override, ?theme= URLs)

Requirements

• A Mac with Apple Silicon (transcription uses mlx-whisper)
• Homebrew: brew install yt-dlp ffmpeg
• Python 3: pip3 install mlx-whisper (the Whisper model, ~1.6 GB, downloads on first use)
• Node.js 20+
• Optional, for the AI features: the Claude Code CLI, logged in (claude → /login). Without it, transcription, library, search, and exports all work — the AI buttons will tell you what's missing.
• Optional: Apple's Command Line Tools (xcode-select --install) build the OCR helper and the native app; Google Chrome, if present, rescues script-rendered articles.

Install

git clone https://github.com/cvdvs/transcript-desk
cd transcript-desk
./Scripts/install-service.sh

That builds the app and installs it as a login service — it's now running at http://localhost:3999, starts with your Mac, and restarts itself if it crashes.

Want it as a real Mac app — Dock icon, own window, no browser?

./Scripts/build-app.sh

Then find Transcript Desk in Spotlight. The app checks the service and revives it if needed, so it always opens to a working screen.

On your phone (same Wi-Fi): http://<your-mac>.local:3999 → share → Add to Home Screen.

What it costs to run

Nothing. Fetching and transcription are local and free. The AI summaries, chapters, translation, and research packs run on whatever Claude subscription you already pay for — summaries use the small fast model and are negligible; research packs use a bigger model and are still only a few percent of a usage window for an hour-long video.

Honest limits

• Apple Silicon Macs only as shipped (mlx-whisper). The pipeline is one file — swapping in another Whisper backend is a contained change.
• yt-dlp vs. the platforms is an eternal arms race — when a site changes, brew upgrade yt-dlp usually fixes it.
• Instagram sometimes requires login cookies for fetching; thumbnails from Instagram expire after a while (cards fall back to a generated tile).
• No login/auth — the app trusts its network. Keep it on localhost, your home network, or a tailnet. Don't port-forward it to the internet.
• Translation defaults to Romanian ⇄ English (the author is Romanian) — translatePrompt in lib/prompts.js is the one thing to edit.

How it's built

Next.js (App Router, plain JS, no database — notes are JSON files), a pipeline that shells out to yt-dlp / mlx_whisper / claude (lib/pipeline.js), and a ~200-line Objective-C + WKWebView native shell (native/main.m) instead of Electron. Whisper runs are serialized machine-wide with a lock; a heartbeat protocol keeps a dev server and the service from stepping on each other's notes; mirror files carry their note's id so two same-titled notes can never overwrite each other.

For AI coding agents

lib/ is the engine: pipeline.js (fetch → transcribe → summarize), store.js (file-backed notes + maintenance sweep), mdsync.js (the Markdown mirror — ownership-checked, never deletes). Dev server: npm run dev on port 4999 (never fights the service on 3999). After code changes: npm run build, then reload the LaunchAgent. The transcription queue, heartbeats (data/.hb/), and the whisper lock (data/.whisper.lock) are cross-process safety — don't remove them.

---

Made by Claudia Vaduvescu (GOODGLYPH) for her own research workflow, and shared as-is. Built with Claude Code. MIT.