X Video Toolkit — one-click video download and local AI transcription

Website · Setup guide · Privacy policy

X Video Toolkit is a privacy-first Chrome/Chromium extension that adds Download and Transcript buttons directly to supported X (Twitter) video posts and Google Drive video previews.

One click can download the best available MP4 or turn the video's speech into copyable text, timestamped notes, TXT, or SRT subtitles. Transcription runs locally on macOS through Spokenly using the Apple-Silicon-optimized Core ML conversion of NVIDIA Parakeet TDT 0.6B v3.

This is currently an unpacked developer extension, not a Chrome Web Store release. It requires macOS, Spokenly, FFmpeg, and a one-time Native Messaging host setup.

What it does

Action	Result
Download an X/Twitter video	Saves the best available MP4 in one click
Transcribe an X/Twitter video	Copies a local AI transcript without uploading audio to this project
Work with Google Drive video previews	Adds the same download and transcript workflow
Copy with timestamps	Produces AI-friendly notes every 10, 30, or 60 seconds
Export subtitles	Creates sentence-level SRT output
Reuse previous work	Stores verified transcripts locally for the configured retention window

Main features

• Inline Download and Transcript actions on supported video pages.
• X/Twitter media resolution for primary, quoted, retweeted, and nested videos.
• The lightest MP4 with the best available audio is preferred for transcription.
• Audio-only HLS and additional MP4 variants remain as fallbacks.
• Bottom-right progress notifications for download, transcription, retry, and clipboard stages.
• Duration validation prevents truncated audio from being accepted.
• Timestamp coverage validation prevents incomplete transcripts from entering history.
• Automatic retries for interrupted downloads and recoverable ASR failures.
• Transcript history with:
  • clean copy;
  • timestamped copy;
  • TXT download;
  • SRT subtitle download;
  • source post/file link.
• Configurable history size, retention, title inclusion, and timestamp frequency.
• No telemetry or hosted backend operated by this repository.

Supported websites and browsers

Websites

• X / Twitter video posts whose media is accessible to the browser/native host.
• Google Drive video preview pages after the active player exposes a media URL.

Protected, deleted, region-blocked, expired, or otherwise inaccessible media may not be downloadable. “One click” describes the interaction, not a bypass of platform permissions.

Browsers

• Google Chrome
• Helium
• Comet
• Arc
• Brave
• Microsoft Edge
• Vivaldi
• Chromium
• Opera

The installer writes Native Messaging manifests for the supported Chromium locations it finds on macOS.

How local transcription works

flowchart LR
    A["X / Twitter or Google Drive page"] --> B["Extension content script"]
    B --> C["Manifest V3 service worker"]
    C --> D["Native Messaging host"]
    D --> E["FFmpeg: temporary 16 kHz mono WAV"]
    E --> F["Spokenly local API"]
    F --> G["FluidAudio + Core ML<br/>NVIDIA Parakeet v3"]
    G --> H["Text + word timestamps"]
    H --> I["Clipboard, history, TXT, SRT"]

Loading

The extension does not run a second 2.5 GB Hugging Face/MLX copy of Parakeet. It calls Spokenly's local transcribe_file API at 127.0.0.1 and verifies that the selected file-transcription model is parakeetTDT06.

Spokenly uses FluidAudio and the Core ML Parakeet v3 conversion, which is designed for Apple hardware. Core ML can use the CPU, GPU, and Apple Neural Engine through the system runtime.

Core ML versus MLX

Both are Apple-Silicon inference technologies, but they are different runtimes and use different model conversions:

• This repository's default: Spokenly → FluidAudio → Core ML → Parakeet v3.
• Possible custom backend: a Python or Swift MLX port with separate weights.

Do not download both unless another application explicitly requires the MLX version. For this extension, the Spokenly/Core ML model is sufficient.

Install on an Apple Silicon Mac

1. Install prerequisites

Install Homebrew if needed, then:

brew install ffmpeg python

Download and install the current sideload macOS version of Spokenly. Keep Spokenly running in the background while using the extension.

2. Download the correct Parakeet model in Spokenly

In Spokenly:

1. Open Settings.
2. Open Dictation Models or Transcribe File, depending on the current UI.
3. Download NVIDIA Parakeet TDT 0.6B v3.
4. Select it as the model used by Transcribe File.
5. Enable Voice for AI Agents (MCP) so Spokenly's local server is available.
6. Optional: enable Local-only mode to block online providers.

Spokenly downloads the optimized Core ML files automatically. You do not need to download the original NVIDIA repository or mlx-community/parakeet-tdt-0.6b-v3 for this setup.

The official NVIDIA model is nvidia/parakeet-tdt-0.6b-v3: a 600-million-parameter multilingual speech-to-text model with punctuation, capitalization, and timestamps. Spokenly uses a converted Core ML build suitable for local inference on Apple Silicon.

3. Clone the extension

git clone https://github.com/Lemelson/x-post-toolkit.git
cd x-post-toolkit

4. Load it as an unpacked extension

1. Open your browser's extensions page:
   • Chrome: chrome://extensions
   • Edge: edge://extensions
   • Brave: brave://extensions
   • or the equivalent page in another Chromium browser.
2. Enable Developer mode.
3. Choose Load unpacked.
4. Select the cloned x-post-toolkit folder.

5. Install the Native Messaging host

From the repository root:

./scripts/install_native_host.sh

The installer normally detects the unpacked extension ID. If detection fails:

./scripts/install_native_host.sh <extension_id>

Then reload the extension and refresh the X/Twitter or Google Drive page.

The installer may create a lightweight parakeet-mlx compatibility command for older local tools. It routes those calls to Spokenly and does not download a second model. If a real parakeet-mlx command already exists, the installer preserves it.

Usage

Download or transcribe from the page

1. Open a supported X/Twitter post or Google Drive video preview.
2. Use the injected Download or Transcript button.
3. Watch the progress notification.
4. The completed transcript is copied and saved to local extension history.

Use the popup

Open the extension popup to:

• act on the current video;
• browse recent transcripts;
• copy clean or timestamped text;
• download TXT or SRT;
• change retention and formatting preferences.

Verified transcripts are reused during the configured retention window, so clicking Transcript again does not rerun the model unnecessarily.

The personal use case behind this integration

This version came from a practical local-AI cleanup:

1. I installed Spokenly and downloaded NVIDIA Parakeet TDT 0.6B v3 inside the app.
2. Spokenly installed an Apple-optimized Core ML/FluidAudio conversion.
3. I asked an AI coding agent to inspect the real local setup and connect this extension to the same model.
4. The agent benchmarked the previous standalone MLX path and the Spokenly path on the same X/Twitter video.
5. After the Spokenly path returned correct text and timestamps, the extension was switched to it and the duplicate 2.5 GB model was removed.
6. A tiny compatibility wrapper kept older local tools working through the same single Spokenly model.

On that machine, one 15.232-second test clip completed in 1.558 seconds end-to-end, including media access and audio extraction; the warmed Spokenly ASR stage took 0.342 seconds. These are informal local measurements, not a controlled vendor benchmark. Hardware, cache state, media delivery, audio length, and background load can change the result substantially. A later cold run after the model had unloaded took 11.2 seconds, so the warmed number is not a startup-time guarantee.

Guide for AI coding agents

If you ask an AI agent to install or adapt this project, give it this checklist:

1. Inspect the real repository, current git diff, and existing Native Messaging host.
2. Confirm Spokenly is installed, running, and using parakeetTDT06 for file transcription.
3. Confirm http://127.0.0.1:51089 exposes the local transcribe_file tool.
4. Run a known audio fixture through Spokenly before changing code.
5. Preserve FFmpeg duration checks and transcript timestamp coverage validation.
6. Never delete an older ASR backend until the replacement passes the same-file test.
7. Search other local projects before removing a shared command or model.
8. Run npm run check, a secret scan, and a real end-to-end transcription.
9. Treat generated code and documentation as untrusted until reviewed.

The implementation contract and JSON examples are documented in docs/AI_INTEGRATION.md.

Privacy and security

• Media still has to be fetched from X/Twitter or Google Drive.
• FFmpeg extracts temporary audio into a per-request temporary directory.
• Temporary audio is deleted after the request completes.
• Spokenly receives a local file path through its loopback-only local API.
• With Parakeet selected, transcription happens locally on the Mac.
• Transcript text, timestamps, source URL, and a short preview are stored in chrome.storage.local.
• This repository contains no analytics SDK, telemetry endpoint, API key, or project-operated transcription server.

See SECURITY.md for the public-repository checklist and responsible disclosure guidance.

Architecture

• content.js — shared page lifecycle and injected UI.
• content_x_adapter.js — X/Twitter post and video context.
• content_drive_adapter.js — Google Drive preview context.
• background.js — queue and message router.
• src/background/context.js — normalized item identity and metadata.
• src/background/media.js — media discovery and audio-quality-aware ranking.
• src/background/native.js — Native Messaging transport and progress stream.
• src/background/store.js — verified transcript cache, history, and settings.
• src/background/format.js — timestamp, TXT, and SRT formatting.
• scripts/native_host.py — FFmpeg, validation, Spokenly API, and result contract.
• scripts/install_native_host.sh — browser host installation.
• scripts/spokenly_parakeet_compat.py — optional legacy CLI bridge.

Development

npm test
npm run check

npm run check validates JavaScript syntax, runs Node and Python tests, compiles the Python host, and checks the shell scripts.

Troubleshooting

Native host is unavailable

./scripts/install_native_host.sh <extension_id>

Reload the extension and refresh the page afterward.

Spokenly is unavailable

• Start Spokenly.
• Enable Voice for AI Agents (MCP).
• Confirm NVIDIA Parakeet TDT 0.6B v3 is selected for file transcription.
• Keep Local-only mode enabled if you do not want fallback to online providers.

FFmpeg is missing

brew install ffmpeg
which ffmpeg

No downloadable MP4 found

The post may be private, deleted, protected, region-restricted, expired, or use a media form not currently handled by the extension.

Incomplete download or transcription

The extension retries recoverable failures. Failed or incomplete transcripts are not cached. Local diagnostics are written to:

~/Library/Logs/xvdl-native-host.log

Uninstall the Native Messaging host

./scripts/uninstall_native_host.sh

The uninstaller removes this project's browser manifests and binaries. It removes the parakeet-mlx compatibility command only when that command belongs to this project.

AI-assisted development disclaimer

This software and parts of its documentation were created and reviewed with the help of artificial intelligence. AI can misunderstand APIs, overlook edge cases, or produce incorrect documentation. Review the source, test with non-sensitive media first, and verify privacy/security assumptions for your own environment.

The project is not affiliated with X Corp., Google, NVIDIA, Apple, FluidInference, or Spokenly. Product names and trademarks belong to their respective owners.

License

MIT