blurt

https://github.com/user-attachments/assets/334087d3-9b56-48fa-808a-2fb229ab666a

Voice alone is blind. Screenshots plus typing is slow. The most natural way to tell an AI what you mean is the way you'd tell a colleague sitting next to you: point at the screen and talk. blurt records exactly that, and your agent does the rest:

• splits a rambling half hour into separate items, even when you jump between topics and correct yourself
• picks the right frame for each one and boxes the exact spot, using where your mouse actually was. Want it exact? Hold ⌃ and circle it on screen
• writes it up: actual vs. expected, repro steps, owner, and the code that's probably responsible
• lets you triage it in a review page, one item at a time with the keyboard
• files it to Feishu/Lark, GitHub, Linear or Markdown, or goes straight to fixing it

What used to take two days of screenshots, red boxes and spreadsheet rows is now a 30-minute walkthrough.

What people blurt

One recording can mix all of these. The agent decides what each item is, and teams can add their own lenses (e.g. ux-research, sales-call, sop).

Get it

npx skills add AGIHunt/blurt

Claude Code: /plugin marketplace add AGIHunt/blurt · or just paste this repo's URL to your agent and ask it to install the skill.

Then, in any project, tell your agent "start blurt" / 「开始口喷」. The first run picks a speech model for your machine and installs the Blurt menu-bar app.

How it feels

1. Draw the area to record: drag a region, click a window, or go full screen. Tabs, bookmarks and everything else stay out of the video.
2. 3-2-1, then talk. A tiny floating bar shows time and mic level, with ⏸ pause, ↺ redo and Finish. The bar itself is never recorded. Shortcuts: ⌥⇧P pause, ⌥⇧S finish. When you say "this", hold ⌃ and drag to circle or underline it (macOS). The ink fades after 3 seconds, and your agent sees exactly what you meant.
3. Click Finish and get back to work. Your agent transcribes locally, writes the items, and opens the review page.
4. Triage like a feed: A keep · X drop · J/K next/prev · Z undo · G list · V overview. Then export, or say "fix them".

Always on: the Blurt app lives in your menu bar. ⌥⇧R starts a recording from anywhere and ⌥⇧R again finishes it; ⌥⇧B opens the menu. Recordings go to a workspace: ~/Blurt by default, or a project you bind, so your agent can process them with the code. Turn on After recording → Claude Code / Codex and every recording gets processed in the background, with the review page popping up when it's ready.

Solo or team

• Solo: record → your agent in the same repo processes and fixes. No forms, no copy-paste.
• Team: teammates without a repo just download Blurt for macOS (unzip, then right-click → Open the first time) and press ⌥⇧R. Anyone records with the app, then uses Recent recordings → Copy video and pastes it into Slack/Feishu. Or bind a shared project folder. Items keep the recorder's name, and exports land in the team's existing tables with your column names.

Under the hood

• Local-first speech recognition. SenseVoice via sherpa-onnx is about 240 MB, very fast on any CPU, and handles mixed Chinese/English well. Whisper on Apple Silicon or NVIDIA. You can also bring your own Groq, OpenAI or DashScope key. By default nothing leaves your machine.
• Native recorder. ScreenCaptureKit on macOS with audio and video in sync to within one frame. Turn on system audio for calls and demos: it goes on its own track, so the transcript tells your words from the other side's. Tk + ffmpeg on Windows.
• Deterministic tools, flexible model. Scripts handle the recording, speech-to-text, frames, the review page and exports. All the judgement is left to your agent (see SKILL.md), so it adapts to your product, your language and your team.
• Any language in, same language out. SenseVoice covers Chinese, English, Japanese, Korean and Cantonese; for German, French, Spanish and the rest (99 languages) use Whisper locally or a cloud key. The first run picks one for the language you speak, and items come back in that language.

FAQ

Does it burn a lot of tokens? The video is never fed to the model. Speech is transcribed locally (free), the agent reads the text, and it looks at a few frames only for the moments that become items. Cost follows how many things you talk about, not how long you record; silence and clicking around cost nothing. Quality follows the model: use one with vision, the stronger the better. Two real sessions (Claude Code, Opus 5.5, a production web app):

Plus cache reads of the ongoing conversation (~4–5M tokens, 1/20 of the input price, included above), which depend on how long your chat already is.

I don't do frontend. Is it for me? Yes, if you can see the problem on screen or film it with a phone. See Not just web apps above.

Roadmap

Browser capture (console errors and network failures lined up with the video) · the pen on Windows · Windows tray app · a hosted speech API · more lenses and exporters · toward a personal assistant that watches, listens and keeps your projects moving. See TODO.md.