Changelog

Every release, newest first.

Echo ships when it's right, and this page says exactly what changed. What's coming next lives on the roadmap.

0.4.1

2026-10-08

Quit finishes a live meeting. Closing Echo in the middle of a meeting now waits for the last pieces and writes the note, instead of leaving a note with no end and no speakers for the next launch to adopt.
The far side's echo on the mic no longer shows as extra speakers. When the laptop mic hears the call through the speakers, that echo is matched against the call audio and dropped, so the other person stays one voice on their own track.
A Meet tab you leave ends the call for Echo too, so the meeting stops when you do.

0.4.0

2026-09-29

Echo takes the notes. A Meetings tab in the Echo window. Start listening from the tray, the tab, or a key you choose, and Echo hears your mic and the other side of the call through your speakers. Stop, and a few seconds later the note is there: a title, a summary, the decisions, the next steps with who owns them, the open questions, and the full transcript with who said what. Dictation keeps working the whole time. Every screen is on the Meetings page.
Who said what. Echo tells the voices in a meeting apart, on your laptop: every line gets a speaker, and a name once you give one. Keep "Remember this voice" on and the next meeting names that person for you. Merge two speakers that are one person, or hide a voice. Echo also guesses names from how people address each other, and asks first.
Your name on your lines. Echo learns your voice from your dictation, so a meeting names you, even a short one.
A meetings library. One search box over every meeting that lands on the line where something was said, with its time. Folders, tags, stars, open to-dos, a filter by person, and a trash you can undo.
Notes with times. Decisions, next steps and topics carry the time they were said; click one to jump to that line in the transcript.
Your notes, expanded. Type what matters while the meeting runs. Echo keeps your lines word for word at the top of the note and fills them in from the transcript.
Notes that fit the meeting. Standup, one on one, sales call, brainstorm, interview, or plain. Pick one during the meeting or afterwards and re-run the note; the transcript is never touched.
Ask your meetings. "What did we decide about pricing?" over one meeting or all of them, answered from the transcripts with the meeting and the time it was said, and saved into the meeting if you want to keep it.
A call comes to the front and Echo offers. Zoom, Teams, Meet, a huddle: the first time a call window shows up, Echo asks whether to take notes. Turn on auto-start in Settings and it just does. No calendar, no sign-in.
Stops when the call does. When the call window closes, or everyone has been quiet a while, Echo stops and writes the note. A dropped call picks up again in the same note.
The room stays out of the call. In a call, voices that are not part of it, like a TV or the kids, stay out of the notes and are one tap away if you want them. A meeting in the room stops by itself after two hours with nobody Echo knows talking.
Speakers, no headset, no double lines. When the laptop mic hears the call through the speakers, the other side's words are kept once, on their own track.
Every note is a file. Plain Markdown in the Echo notes folder, with names, people, folder, tags and checkbox action items, plus an optional second folder for a project or a synced drive, so any tool that reads a folder has the context.
Cleaner meeting transcripts. Meeting audio no longer goes to the speech model with your dictation word list, which turned some long, clear passages into "and so on", quiet meeting audio is brought up to level before it is sent, and a voice across the room is no longer skipped as silence.
Settings, rebuilt. Settings opens as a window over the app, with General, Dictation, Meetings, Models, Keys, Usage, Privacy, Appearance, Account and About.
See your keys. Keys shows every API key with its state (working, rate limited, out of quota today, needs billing, key rejected, model retired, slow), when it last answered and how fast, and the requests left today where the provider says. When a key or model stops working, a dot shows on your avatar and Echo says so once.
See your usage. A 30-day chart of model calls, tokens and speech minutes, with dictation and meetings apart, today's use per key, and your lifetime totals since your first dictation.
Pick your models. Models shows the order Echo tries for each job and lets you pick the main cleanup model.
More ways to keep going. Cleanup tries three models on every key before the next provider, with a time limit on each. A meeting too big for one request goes up in parts. Rewrites go through the same fallbacks as dictation. With a Cloudflare Workers AI key, speech to text falls back to Cloudflare before the offline model.
Light mode for the Echo window, or follow the Windows setting.
The pill shows a Notes clock while Echo listens, and goes back to it after every dictation in the middle of a meeting.
Pasted text uses straight quotes and plain hyphens, never curly quotes or long dashes.
A very short, quiet dictation can no longer paste Echo's own word list.

0.3.7

2026-09-27

A long dictation survives a dropped connection. The speech service sometimes drops the upload on clips longer than about 15 seconds, and Echo used to give up within a second and ask for a restart. Now it retries on a fresh connection, then on a spare key, then transcribes on your computer with the offline model when you have it installed. If all of that fails, the audio is already safe on disk, Echo keeps retrying in the background, and when the words come back they land on Ctrl+Alt+V with a notification. No restart needed.
Echo stopped turning "me" into "em". A learned correction never rewrites a common word anymore. Name fixes built from common words still work, and the History edit tells you when a fix is too common to apply.
Cleanup can no longer delete what you said. If the cleaned text loses more than about 15% of your spoken words, Echo pastes the lighter cleanup instead.
Instructions to AI agents stay as you said them. In Cursor, VS Code and terminals, "make the title all caps in blue" stays words for the agent instead of being carried out on your text. Spoken punctuation still works, a sentence you cut off no longer gets a period added, and the next dictation into the same box still starts with a space.
Names come out right more often. The vocabulary Echo sends to the speech model now fits the service's size limit, with your own names first instead of generic tech words. Words you add in the Dictionary go to the top of the list.
Your screen stays where you are. Move to another window while Echo works and the text still lands where you started talking, then the window you moved to comes back in front. Switchable off.
Hands-free warns you and stops at 5 minutes. A notification at 4:30. At 5:00 the words go to the Scratchpad and Ctrl+Alt+V instead of into the window you started in. Before, everything past 5 minutes was silently dropped.
Cleanup uses the style of the window you started in, and History records that app, even if you were looking at another window when you let go.
Fix last goes back to the window the last dictation landed in.
Each dictation reads only the corrections it needs, about 50 ms faster.

0.3.6

2026-09-21

The pill tells the truth about every paste. Pasted when the target still had focus after the keystroke, Pasted (unverified) when the keystroke went out but focus had moved, and On clipboard when Echo held the text back rather than paste it into the wrong window. That last one stays up long enough to read and raises a notification, so a dictation that did not land can no longer look like one that did.
Paste the last dictation again with Ctrl+Alt+V, or from the tray menu. This is the recovery for a paste that did not land, and every path that re-pastes now uses the chord the target window actually wants.
Terminals get Shift+Insert. Windows Terminal, the classic console, mintty, Alacritty and WezTerm receive Shift+Insert instead of Ctrl+V, which the Claude Code prompt was quietly dropping. Any app can be put on either chord from your .env.
Dictations into Cursor stop vanishing. Echo no longer sends a stray Alt release before pasting; Chromium apps read that as open the menu bar and move focus off the text box. When Echo has to bring the target back to the front it now waits 150 ms and checks afterwards that the target kept focus.
A dead cleanup provider is set aside for the session instead of being retried on every dictation. Polish that took 3 to 7 seconds with dead keys in the chain is back to about one second, and a single long dictation can no longer knock a working provider out for the rest of the day.
Groq deleted its Llama models. If your settings still name one, Echo swaps in the current model and says so, instead of failing every cleanup in silence.
History shows how long your dictations take: transcribe, cleanup, insert and release to text, with the median and the 95th percentile over your last 200 dictations.
Echo brings its own window back after a paste when you had clicked into it while the cleanup ran, instead of leaving it buried behind a maximized app. Switchable off.
Dictation history is kept forever instead of being trimmed to 30 days, and the app window remembers its size.
Echo's clipboard restore is skipped when something else wrote to the clipboard in between, so a clipboard manager no longer gets the old contents back from under it.
Dictating into the Echo window sends the text to the Scratchpad tab and says so.
A hotkey that is only a letter or a digit is refused, because the letter would be typed into whatever you were writing.

0.3.4

2026-09-13

No more stuck Windows key. A fast Ctrl+Win release could leave Windows thinking the Win key was still held, so every letter after it fired a Win shortcut until a reboot. Echo now keeps the Win key fully hidden from Windows for the whole hold, and if Windows saw it anyway, hands it the matching release.

0.3.3

2026-08-22

All-time means all time. The voice profile's lifetime word count was quietly showing only the last 30 days. It now reads the same lifetime ledger the dashboard uses, and dictations from June 2026 are back in the totals.

0.3.2

2026-08-21

See your words as Echo hears them. The pill shows raw speech on top and the cleaned text below, so you can see exactly what Echo changed.
Fix the last paste without speaking again. A Fix last key undoes the paste, lets you edit the text, and pastes it back. The correction also teaches Echo, so the same mishear stops coming back.
Short dictations get your learned corrections too, not just the long ones. Transforms now use your dictionary, your corrections, and the app you are writing in.
Transform hotkeys can be switched on from Settings instead of the .env file.
After each paste the pill shows that dictation's word count and pace.
Checking for updates shows a small one-line status on the pill instead of the full dictation pill with dead bars.
Pasting into terminal panes that use the Kitty keyboard protocol works now. Echo sends the paste chord at human speed instead of instantly, which those terminals were silently dropping.
If the overlay pill wedges on a cold boot, Echo gives up on that attempt at 20 seconds instead of a full minute.

0.3.1

2026-08-07

Updates stay on screen the whole way. The window stays up until the installer takes over, progress is visible, and Echo restarts itself when it is done. If anything goes wrong, the dialog says so.
A What's new panel after every update, with the roadmap on a second tab.
The overlay pill spells out update checks in words: Checking for updates, Up to date, Update check failed.
Echo opens its window when you launch it, and launching it a second time raises the existing window.
The installer puts the WebView2 runtime on machines that do not have it, so Windows 10 installs work out of the box.
Very quiet clips are treated as silence instead of being amplified into invented words.
Fixed a startup hang that could keep the overlay pill from ever appearing, and an invisible menu that swallowed clicks in the app window.

0.3.0

2026-07-15

Hands-free dictation that actually works. Double-tap the push-to-talk key to lock the mic, tap once to stop. A slow audio device can no longer eat your double-tap.
Your text lands where you started talking. The paste target is frozen the moment you engage a dictation, so moving windows mid-sentence never sends text somewhere else.
Audio rescue. A dropped connection can no longer eat your words. If transcription fails, the recording is kept on disk and transcribed on the next launch, then deleted. Local only, and you can switch it off.
A real dashboard. Your median words per minute against typist percentiles, hours saved measured against the minutes you actually spent talking, a 24-hour rhythm chart, an activity heatmap with streaks, and personal records.
Lifetime stats stay lifetime instead of quietly becoming 30-day stats.
Transcription requests time out instead of hanging the pipeline.
The Rewrite hotkey defaults to Right Alt, a key every laptop has.
No more surprise Start menu mid-dictation. Two phantom Windows key bugs fixed at the root.
Heavy screen updates wait for a lull and never run while the mic is hot.
The pill has an honest hands-free state: live coral pulse, labeled Hands-free.

0.2.0

2026-07-12

One overlay. The pill relaunches itself if it cannot start, and dictation keeps working without a visual rather than stalling. The overlay can never take dictation down with it.
Privacy controls. Pause dictation from the tray for 10 or 60 minutes, and keep a per-app block list so Echo never records over a password manager or a banking window.
Config backup. Export your dictionary, snippets, profiles, modes, and settings as one zip and import on another machine. API keys are stripped on the way out and refused on the way in.
New Transforms tab. Run Polish, Concise, Formal, Bullets, or Developer on your last dictation.
New Modes editor. Change how each cleanup style rewrites you, add your own, and try them on sample text first.
The learning loop pays off: Echo tracks when a correction it learned actually fixed a live dictation, and the Learned tab shows which corrections are earning their keep.
Silence hallucinations are filtered at the source. Phantom phrases out of a quiet moment are gone.
Lists form from real speech, and hold together across an interrupted push-to-talk burst.
Cleanup knows what you are typing into. Terminals keep filenames, flags, and tokens intact.
The clipboard survives a dictation: whatever you had copied is restored after Echo pastes.
A first-run wizard, a Scratchpad tab, sectioned Settings, a word-level diff view in History, and a full visual pass across the app.

0.1.9

2026-06-18

Push-to-talk accepts key chords, not just single keys, and defaults to Ctrl+Win.
Pick which microphone Echo records from.
The overlay pill re-pins itself on top, so other windows stop covering it.

0.1.8

2026-06-13

Fixed an install that could fail outright when the desktop shortcut box was ticked.
The overlay pill no longer shows up as its own taskbar button.
The correct Echo icon shows immediately after install, and uninstalling removes the desktop shortcut.

0.1.7

2026-06-13

Fixed the in-app updater, which was silently broken. Clicking Install update now installs the update.
The installer waits until Echo has actually exited before replacing files.

0.1.6

2026-06-13

A real Settings page: speech model, language, cleanup model, and update behavior are live controls.
Paste no longer follows your focus into the wrong window. If focus leaves the target, the paste is cancelled rather than dumped somewhere else.
A dead microphone gets an honest message and a real retry.
Dozens of smaller reliability fixes across Settings, paste, and the microphone path.

0.1.5

2026-06-12

Dictate without your own API keys. Access rides along with your invite, and the keys never leave the server.
The learning loop. Echo notices words it keeps mishearing, groups them by where you were writing, and offers corrections as chips. Confirmed corrections stop the same mistake from coming back.
Editing a dictation in History teaches Echo: a fixed word becomes a correction, a reshaped sentence becomes an example of how you write.
Writing surfaces: vocabulary, substitutions, and snippets can be scoped to a specific app.
Rapids, the new default waveform, plus Bars, Studio, and Cloud in a new Appearance tab.
Double-tap to lock the mic hands-free. Smart spacing between consecutive dictations. Day-grouped History with full-text search.

0.1.4

2026-06-11

Check for updates works from the tray, and a found update opens the install dialog directly.
Cleanup knows which app you are writing in and adjusts tone and formatting to match.
The compact pill becomes the default overlay, and the cleanup chain survives a quota day by falling through to the next provider.

0.1.3

2026-06-10

Auto-update runs off the Echo release feed, verified by checksum, and refuses an install if the download does not match.
The Echo desktop window: Home, History, Snippets, Dictionary, and Profiles, live as you dictate.
Hallucination gates: phantom phrases from silence are caught and dropped rather than pasted, and a punctuation guard keeps a stray period from submitting a terminal command.
Dictionary suggestions: Echo spots proper nouns it keeps butchering and offers ranked fixes.
Cleanup that fails loudly. A rewrite that errors says so instead of pasting the wrong thing.

0.1.2

2026-06-08

Multi-provider cleanup chain. If the primary service is down or rate limited, Echo falls through spare keys, then a second provider, then a local rules pass, rather than handing you raw transcript.
An always-on compact pill overlay, and auto-update off the Echo release feed.
The speech model is biased with your vocabulary, so your names and jargon come through.
Spelling mode and self-correction: say it wrong, correct yourself out loud, and only the corrected version gets pasted.

What's next: the Echo roadmap.