Open Source Radar — November 6: VibeVoice's Hour-Long Transcript
Today's trending board is heavy on tooling rather than new models: Microsoft's speech family is back, two agent-facing CLIs are shipping daily fixes, and Google has quietly built a 187-skill catalog for its own cloud.
VibeVoice (Python, about 56,300 stars, MIT) is Microsoft's open speech family, and the one most people will actually reach for is VibeVoice-ASR, which accepts a full 60 minutes of audio in one pass inside a 64K-token window and returns a structured transcript — who spoke, when, and what was said — rather than a wall of unlabelled text. It runs in more than 50 languages, takes custom hotwords so it stops mangling names and jargon, and ships finetuning code and vLLM serving. The detail that makes it practical on a laptop is the separate BitNet build: the model drops from 4.62 GB to 1.58 GB and still transcribes faster than real time on three CPU threads, with no GPU in the machine. One thing to know before you clone it — Microsoft pulled the text-to-speech code in September 2025 after finding it used in ways it did not intend, so this is a transcription project now, and the TTS demo in the model table is disabled.
agent-browser (Rust, about 46,200 stars, Apache-2.0) gives an agent a browser it drives one shell command at a time, with no screenshot loop and no vision model in the path. A snapshot returns the page as compact text where every actionable element carries a ref, the agent clicks the ref, then reads again to confirm; the browser stays open between calls because each command goes to a background daemon over the Chrome DevTools protocol. The detail worth stealing is the failure behaviour: a click reports an error when the target is hidden, disabled, off-screen or covered by a consent banner, instead of pretending it worked and letting the agent build on a lie. It also ships its own version-matched skills, an MCP server, and a choice of engines including Lightpanda and Obscura — committed to as recently as yesterday morning, which is why it keeps surfacing.
google/skills (Python, about 22,600 stars, Apache-2.0) is Google's own answer to the skills format, and the size is the story: 187 skill folders covering IAM onboarding, BigQuery AI, the Gemini API, Genkit in four languages, GKE inference migration, agent gateway security and a long tail of Cloud Storage and Filestore operations, plus a separate group for the Ads API. Commits land most weekdays — yesterday added an API Gateway troubleshooting skill and usage attribution in the bundled MCP configs — and the whole set installs into Claude Code, Codex, Cursor or Copilot through the skills registry rather than one vendor's plugin format. Read it as documentation shipped as skills: it teaches an agent to do things the Google way, and several of those skills will call billable Cloud APIs the moment an agent follows them, so scope the service account before you hand it over.
ai-goofish-monitor (Python, about 18,100 stars, MIT) is a monitor for Xianyu, Alibaba's second-hand marketplace, and it is the most concrete "agent that watches something for you" project on the board today: Playwright scrapes listings for your keywords, a multimodal model reads the photos and description against criteria you write in plain language, and anything that passes gets pushed to ntfy, WeChat Work, Bark, Telegram or a webhook. It keeps tasks, results and price history in one SQLite database, filters by free shipping, listing age and province-to-district region, and runs on a cron schedule behind a web admin UI, with Docker for deployment. The honest caveat is the account and proxy rotation — the feature exists because marketplace logins get rate-limited — so run it on an account you are willing to lose. It trended today with no code pushed since October 10, which says more about demand for deal-hunting agents than about the release.
Worth watching this week: whether Google keeps the daily commit pace on its skills repo, and whether VibeVoice's ASR line gets a streaming release that is not labelled research-only.
What would you want an agent watching for you — prices, job listings, or something stranger? Write to us via our contact page.
