Blurt is a menu-bar app for the Mac, with test builds for Windows and Linux. It works next to the agents you already use, such as Claude Code, Codex, Cursor,
Claude and ChatGPT, and handles the part before they start: getting what you mean into the agent's input box. It has
no agent of its own and doesn't replace yours.
What it does
Dictation (Fn). Tap Fn in any app, talk, and tap it again. The words are typed where your cursor is. Finish
with a double tap instead and it's sent too (Return). It works the moment Blurt is installed, with macOS's own speech
recognition (macOS 26) and nothing to download. Open-source local models are an optional upgrade: more accurate, they
learn your words and need no internet; after a minute of talking the text is in about 0.4 s. A local model removes
fillers and stutters and adds punctuation; it doesn't rewrite or summarize. The words show under the notch as you
talk. Double-tap ⇧ while talking to switch to Structured (a list, when you listed things) and again to switch back;
double-tap ⌘ to translate (speak Chinese, type English, or the other way round); ⌥L reformats the last dictation. Holding Fn
while you talk works too.
Snap (⌥A) and Voice Note (⌥V). Drag over part of the screen and say what's wrong; the note opens right
beside the selection. Draw on it as you talk: drag for a freehand line, ⇧ for a box, ⌥ for an ellipse, ⌘Z to
undo. The snap becomes a card under the notch. Drag the card into Claude, ChatGPT, Slack or a terminal running Claude Code or Codex: the picture and
your note go in together. Or click it to copy, then ⌘V. A voice note is the same, with your words and no picture.
Copy text from screen (⌥C). Drag over any text you can't select: a screenshot, a video, a locked page, a
terminal. Let go and it's on the clipboard in about 0.2 s, laid out as if you had copied it: the half line the edge
cut is left out, wrapped list items are one line again, tables keep their tabs and code its indentation. It reads with
macOS's own recognizer, so there's nothing to download. Optionally a second reading by a small local OCR model
(720 MB) or your own cloud model fixes misread symbols in code and paths.
Record (⌥R). Record your screen while you talk through a product, a page or an idea. Your coding agent turns
the recording into items (issues, ideas, notes, tasks), each with a screenshot, the spot marked, steps and the code
that is probably involved. It's transcribed while you record, so it's done about a minute after you stop; one
5-minute walkthrough became 52 items, each with what was said and when. You go through them with the keyboard, then
hand them to an agent, or pack them into one .zip for a teammate.
Blurt on your phone. When you sit back from a big screen, your phone becomes the computer's trackpad, keys and
microphone, and can show its screen. It works with a Mac, Windows or Linux computer, from an iPhone, iPad (including the
iPhone Duo) or Android phone or tablet. The talk key is Blurt's dictation key, and the keys follow the computer you're
driving (⌘ on a Mac, Ctrl and Win elsewhere). With several displays the picture follows the pointer, or stays on the one
you pick. Away from home it goes through a relay you run yourself; everything on the way is end-to-end encrypted. Pair
in Settings → Phone on any computer. The phone apps aren't in the stores yet: build them
yourself (iPhone and iPad, Android).
Snaps, voice notes and recordings go into the current project (a folder you open in Blurt). Without one, recordings
go to Unassigned and snaps stay as cards until you pick a project. Every shortcut can be changed in Settings.
Install
You need macOS 13 or later. The local models need Apple silicon; an Intel Mac dictates with macOS 26's own speech
recognition or a cloud key. Windows and Linux: see the FAQ.
Download Blurt for Mac (signed and
notarized), open the disk image and drag Blurt into Applications. Updates arrive inside the app. There is nothing
else to install: the app fetches the Python runtime its tools use (about 150 MB) by itself the first time.
Or give this to your agent (Claude Code, Codex…), or paste it into a terminal:
curl -fsSL https://blurt.work/install.sh | sh
It downloads the latest release, checks that it is signed and notarized, puts Blurt in Applications and opens it.
Use it from your agent. In Blurt, Settings → Agents → Connect… adds the Blurt skill and MCP server to Claude
Code, Codex and the other agents it finds (Gemini CLI, Kimi Code, OpenCode, Grok, CodeBuddy Code, WorkBuddy,
Cursor). Start a new session of the agent, then tell it "start blurt" (/blurt in Claude Code, $blurt in
Codex), or ask it about your items. Your agent can also do the connecting: llms.txt has
the steps.
Without the app, from source: npx skills add AGIHunt/blurt -g, or in Claude Code /plugin
marketplace add AGIHunt/blurt and then /plugin install blurt@blurt. Then "start blurt". The first run
builds Blurt into ~/Applications; it needs the Xcode Command Line Tools and
uv, and the agent offers to install what's missing.
First minute. The setup assistant asks for Microphone and Accessibility (and Input Monitoring if you use another
key than Fn). Then press Fn in any text field: on macOS 26 dictation already works with macOS's own recognizer,
while your Mac's own local model downloads in the background, once (1.6–5.9 GB depending on the Mac). On macOS 13–15
you download a model first or add a cloud key. One optional upgrade: an AI-service key (fixes misheard names and
punctuation, translates with the cloud model, and gives copied text a cloud second reading). Screen Recording is asked
for the first time you snap or record.
How it works with your agents
Blurt gets your words, pictures and recordings into an agent. The agent does the work.
以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗
