Open source · Apache-2.0
中文 ↗

AutoWeft

The open-source browser AI agent: ask about any page, extract data — and, uniquely, transcribe video.

An AI workbench in your sidebar: reads the page, answers questions, structures data into tables — and transcribes video to a transcript in one line. Local-first, your keys, anonymous opt-out telemetry.

★ GitHub

Real screen recording (61s): say “transcribe this video” on a Bilibili page → media detected → approve the extraction → local Whisper transcription → timestamped transcript, all in the sidebar.

Setup after install

Two minutes in the sidebar settings — bring your own keys, or go fully local.

1 · Configure the chat model

Open the sidebar → Settings → Models:

  1. Pick a provider preset — OpenAI-compatible or Anthropic-compatible.
  2. Paste your API key (stored locally, sent only to the endpoint you chose).
  3. Self-hosted / local endpoints (Ollama, vLLM…) work too — any OpenAI-compatible URL.

2 · Speech recognition (ASR)

local option

Sidebar → Settings → Models → Speech recognition (ASR):

  • Cloud presets — Groq (free tier), SiliconFlow, or OpenAI; paste the key, done.
  • Local model — switch the provider to Local model, pick a Whisper size and hit Download: base ≈80MB (default, recommended) · tiny ≈40MB (fastest) · small ≈250MB (best quality).
  • Local runs entirely in your browser (transformers.js) — audio never leaves the device, no key needed.

Then just say “transcribe this video” on any supported page. Full docs: GitHub README

What it can do

Open any page and talk in the sidebar — read, answer, transform, store. All local-first.

💬 Page Q&A & analysis

Answers grounded in the live page: summarize long articles, explain content, compare options. Reads first, never makes things up.

📊 Data extraction & processing

Structure page content into tables (CSV download) and markdown notes — searchable and reusable.

🎬 Video transcription

unique among open source

MSE sites like Bilibili and Douyin, plus direct links and HLS — say "transcribe this video": the audio track is detected, extracted locally, transcribed with resumable ASR into a timestamped markdown transcript.

📄 Full-page export & archive

Export the whole page as markdown (images localized); data, transcripts, and notes live in one searchable Data tab.

$0
local transcription · runs in your browser
33
gated capabilities · per-action authorization
0
content servers / accounts · anonymous opt-out telemetry
80MB+
local ASR model · download once, transcribe offline

Compared with open-source peers

AutoWeft Nanobrowser browser-use
Page Q&A / understanding ✓ read-before-answer, verified context ✓✓
Video/audio transcription ✓ local extraction + resumable ASR ✗✗
Full-page markdown export (localized images) ✓ ✗✗
Tables → Data tab (CSV) + markdown notes ✓ ✗✗
Per-action network authorization (tool+target bound) ✓ global or none—
Local-first (media never leaves device) ✓ partialmostly cloud browsers
Task reliability (plans / resume / breaker) ✓ basicbasic
Form factor Chrome side-panel extension Chrome extensionPython/TS library

Comparison based on public features as of 2026-09; file an issue if anything is off.

✓ Anonymous telemetry, one-tap off ✓ No servers ✓ No accounts

Media is processed in the browser; agent-initiated network requests need per-action approval. Privacy

Install / Update

No Chrome Web Store yet — the installer writes the build straight into a folder you pick.

Latest:loading…   Installed:unknown

· Updates reuse the published file hashes. Unchanged files are checked by size and modification time without reading their contents. Use full check if files appear damaged.

About 20 MB to download · 65 MB unpacked · File purposes

The current download is about 20 MB, or 65 MB unpacked. The main files are:

  • Audio processing engine: about 32.2 MB, for extracting audio tracks and converting formats.
  • Local speech recognition runtime: about 26.9 MB, for running recognition in your browser.
  • PDF character maps: 168 files totaling about 1.2 MB, for text extraction including CJK.
  • Other files: about 4.3 MB of interface, task logic, PDF parsing, fonts and licenses.

Sizes above are unpacked. Speech model weights are not included; download them separately in extension settings when needed. After unpacking, you can inspect every file’s actual size and purpose.


          
First-install guide
  1. Click ① to create or pick an empty folder (e.g. autoweft-ext).
  2. Click ② — the build is written into that folder.
  3. Click ③ and follow the dialog to load or refresh on chrome://extensions.
  4. For future updates, repeat ②③; Chrome usually auto-reloads after files are overwritten.

Chrome / Edge only (requires the File System Access API).

③ Reload the extension

Web pages cannot navigate to chrome:// URLs — copy the address below and paste it into the address bar:

chrome://extensions/

First install (not loaded yet)

  1. Turn on Developer mode (top right of the extensions page).
  2. Click "Load unpacked" and pick the folder from step ①.

Update (already loaded)

  1. Find the extension you loaded on the extensions page.
  2. Click ↻ (reload); Chrome usually auto-reloads after step ② anyway.