Blurt

The keyboard for the agent era.

The fastest way to work with your agents, by text, voice, screenshot or screen recording. Runs locally. Free and open source.

Download for Mac

Free, macOS 13 or later. Windows and Linux are coming soon.

Tap fn and talk. It types where your cursor is.

  1. Tap fn and start talking
  2. Your words show under the notch as you talk. Grey words may still change.
  3. Double-tap fn when you finish. It types the text and sends it.

Works in any app. In a terminal the text goes in as one line. You can also hold fn while you talk and let go to stop.

Done in 0.39 s. No internet needed. It runs on your Mac.

Where speech is recognised
Your Mac
The other two: cloud
Price
Free, no word limit
Typeless Pro from $12/month
Talk for a minute, stop, until it’s done

Blurt

0.39 s

Wispr Flow

≈1.5 s

Typeless

≈3.0 s

Blurt: median of 3 takes on our test Mac. Wispr Flow and Typeless figures: VoiceDash

Compared with four dictation apps

BlurtTypelessWispr FlowSuperwhisperAqua Voice
Where speech is recognisedYour MacCloudCloudLocal or cloudCloud
PriceFreeno word limitWeekly limitPro from $12/mo2,000 words a weekPro from $12/moFree with local modelsPro $8.49/mo1,000 words freePro from $10/mo
Stop talking to done0.39 s≈3.0 s≈1.5 sNot published≈0.45 s
Snap and voice cards
Recordings into items

Blurt’s recognition error rate, lower is better

2.7%
Chinese and mixed Chinese–EnglishCharacter error rate (CER), 20 clips with product names, 90–97% of the names right
1.9%
EnglishWord error rate (WER), the speech model alone
All accuracy figures and how we measured
Blurt’s accuracy
CER 2.7%Chinese and mixed Chinese–English, 20 clips with product names. 90–97% of the names right.
CER 26.7%The same model on a harder set, 6 voices with room noise added. 38% of the names right.
WER 1.9%English, the speech model alone. 7 product-feedback sentences, 2 voices each.
CER 10.8%16–24 GB Macs, the 0.6B model with polish, Chinese and mixed. 86% of names right (measured on v0.8).

None of the other four publish Chinese figures. English: Wispr Flow about 3.9% and Typeless about 7.9% (the same article also says about 20%), from VoiceDash. Aqua Voice publishes an average WER of 6.24% on the OpenASR leaderboard.

Blurt’s numbers were measured on one Apple M5 Max (128 GB), with speech made by macOS say as the input. That is cleaner than a real microphone, so the error rates are optimistic. Real microphones, accents and smaller Macs were not measured. 0.39 s is after 60 s of speech, the median of 3 takes; 42–54 s takes with live text on took 0.51–1.55 s, median 0.64 s. Prices and policies checked on 2026-10-25.

Tap Shift while you talk and get a list

  1. By default it removes the ums, adds punctuation and keeps your words
  2. ⌥L turns the last one into a list. Undo changes it back.
  3. Tap ⇧ while you talk to get the list straight away. If you correct yourself, only the correction stays.

Book the flight for Monday, I mean Tuesday morning.Book the flight for Tuesday morning.

Talk in Chinese. It types English.

  1. Tap fn, talk in Chinese
  2. ⌘ Tap ⌘ while you talk
  3. Stop. English in about 1 s.
  4. Talk in English, get Chinese

With an AI key, the cloud model translates in about 1 s. Without one, the model on your Mac does it in about 0.2 s, a little rougher. Besides Chinese ↔ English it translates into Japanese, Korean, French, German, Spanish and more, best with the cloud model.

Snap. Talk. Drag.

  1. ⌥A Drag over an area. That part of the screen freezes.
  2. A note opens beside the snap. Say what’s wrong and the words show up in it.
  3. Draw on the snap while you talk
  4. Stop for 2 seconds and it’s saved. The snap goes up to the notch as a card.
  5. Drag it into Claude. The picture becomes an attachment and your words go in the message box.
  • Box
  • ⌥Ellipse
  • DragFreehand
  • Rest 0.5 sStraighten
  • ⌘ZUndo
Chat, docs
Picture and words in one message
Terminals
File paths only; nothing runs
Cursor
The picture; words on the clipboard
Voice note
⌥V talk only, up to 10 minutes

Copy text you can’t select

  1. ⌥C Drag over text you can’t select
  2. Let go. Copied in 0.2 s.
  3. Paste it, layout and all
  4. Plain OCR: broken lines and garbage included
  5. ⌥T Translates it into Chinese

Wrong: characters misread on 18 test crops. Plain OCR is macOS’s own recognition, as is. Blurt uses the same recognition, so there’s nothing to download. An optional second reading (a 720 MB local model, or your cloud model) gets it to 0.16%.

Record while you talk. Your agent writes the items.

  1. ⌥R Pick a window or an area, then use it and talk
  2. ⌃ Hold ⌃ and drag to circle a problem
  3. It transcribes while you record, so it’s ready about a minute after you stop1
  4. Your coding agent writes it up as items, each with a frame, the steps and the code that’s likely at fault

52items

From one 5-minute walkthrough, each with what was said and when

The frame at 0:06: Orbit’s checkout in Canada, the total circled with ⌃0:06

OR-21Issue

Total ignores coupon in Canada

“And the total ignores the coupon when I pick Canada.”

  1. Apply SAVE10: Total $117.10
  2. Change the country to Canada
  3. SAVE10 still says Applied; Total is $128.40

ActualSAVE10 shows Applied, but nothing comes off: $128.40.

ExpectedSAVE10 takes 10% off the items ($11.30): $117.10.

web/src/cart/reducer.ts:88 recalcTotals()

1 Measured once for the README FAQ: a 5-minute, 12-item recording was ready to review 51 s after Finish. Organising uses your own agent; three real 7–20 minute sessions cost about $2–2.6 each at API prices.

Hand it to an agent. Merge after you check.

  1. ⌘↩ Select items and press ⌘↩. Claude Code or Codex starts on them, one git worktree each if you like.
  2. When the agent says it’s fixed, Blurt only marks it To verify
  3. You look over the change, then merge

Pack it into a ZIP. Hand it to a teammate.

  1. Pack the items into one .zip for a teammate. It has a summary, every item and the screenshots.
  2. Or ask your agent to “bundle the checkout issues and post them to #web in Slack”

MCPAny agent can read and update your itemssearch_items get_item update_item mark_fixed bundle_items

No account. Your data stays on your Mac.

All on your Mac

  • Speech recognition and polish
  • About 500 characters near the cursor, for names; never password fields
  • Dictation audio and text, kept 30 days; you can turn it off
  • Snaps, voice notes and recordings
  • Copying text from the screen

Goes online only for

  • Model downloads and update checks
  • API key mode: requests go to your provider
  • Send Feedback: only the dictations you check
From the film: Peep in the notch next to the name Blurt Watch the film2:02

Blurt in two minutes

  1. Dictate0:07
  2. Speed0:25
  3. No download0:30
  4. Lists0:34
  5. Translate0:43
  6. Snaps, cards0:55
  7. Copy text1:10
  8. Record1:21
  9. Hand off1:38
  10. Teams1:49

Works out of the box. Your choice of models.

0MB

of models to download

  1. It uses the speech recognition built into macOS
  2. Tap fn, say something, it types
  3. Optional: an AI key for accuracy and translation, local models for offline use

Needs macOS 26. If macOS doesn’t have its speech files yet, it downloads them once (about 400 MB). On macOS 13–15, download a local model first.

Talk to your agents with Blurt

macOS 13+
Macs with Apple silicon
0 MB
to download before you dictate (macOS 26)
30
languages; Chinese and English tested
MIT
Open source on GitHub
Requirements, models and permissions
System
macOS 13 or later. Dictation and the recommended speech models need Apple silicon. Windows and Linux are coming soon. No phone app.
Models
On macOS 26, Blurt uses macOS’s own speech and text recognition, so there’s nothing to download. Local models are optional: Qwen3-ASR 0.6B with Gemma 4 E2B on Macs with 16 GB or more, about 4.4 GB to download, about 2.8 GB of memory in use. Nothing downloads until you press Download. Idle models unload after 10 minutes, sooner when memory runs short.
Languages
Dictation was tested in Chinese, English and the two mixed. The speech model lists 30 languages; the others weren’t measured. The app is in English and Simplified Chinese.
Words
Add project names, libraries and teammates. A word you fix twice is learned, with Undo. Words are hints, never replacements.
History
Every dictation stays in Dictation History for 30 days, with Copy. The mic opens on the key press, so the first word isn’t cut off.
Permissions
Microphone; Accessibility, to paste and hear fn; Input Monitoring, for the other keys; Screen Recording, asked the first time you snap, copy text or record
Typeless or Wispr Flow installed
They use fn too. Blurt tells you when it sees one, so you can quit it or move Blurt to Right Option.

The full user guide

Download for MacFree, macOS 13 or later

Or have your agent install it:

curl -fsSL https://blurt.work/install.sh | sh

Give it to Claude Code or Codex, or paste it into a terminal. It downloads Blurt, puts it in Applications and opens it. You turn on the microphone and Accessibility yourself, in the setup window.

To use Blurt from your agent, connect it in Blurt’s Settings → Agents, start a new session, and say “start blurt”.

Eight keys to know

  • fnTap to talk, tap again to stop
  • fnfnDouble-tap: stop and send
  • While talking: make it a list
  • ⌘While talking: translate
  • ⌥ASnap, draw on it, say it
  • ⌥VVoice note
  • ⌥RRecord
  • ⌥CCopy text from screen (⌥T: translate)

The other keys (re-format, pause, review…) are in the user guide. You can change each one in Settings.