What shipped. What is next

Changelog

Major public-facing updates in the current Echo Flow launch cycle.

October 30, 20261.7.9 · build 179

Echo Flow 1.7.9, the permission prompt stops coming back

• Fixed: the repeated permission prompt. Echo Flow would occasionally ask for Device Control and Data Access (what macOS 27 calls Accessibility) even though the setting was already on, at launch and again on some dictations. Asking was possible from the launch check and from every dictation attempt rather than only from setup, so an app that could not hold the grant re-asked indefinitely. macOS also ties this permission to the app's code signature, so a development build could never inherit the grant the installed app already had.
• The system prompt is now shown once, from an explicit click during setup.
• Everywhere else Echo Flow checks quietly. Launch, dictation, Clinical Note and Quick Rewrite no longer raise a system alert. If access is missing you get a clear message with a button that opens the right Settings pane.
• The guidance names the current pane. macOS 27 renamed "Accessibility" to "Device Control and Data Access"; the message now says so. If you are already granted, nothing changes.
• No model re-download.

October 28, 20261.7.8 · build 178

Echo Flow 1.7.8, sooner live text and replacements that work

• Live text appears sooner. The streaming decoder waited for 1.5 seconds of new audio before its first decode. On a real machine, first live text arrived at 1.94 seconds and that figure was identical for a four-word and a thirty-one-word utterance, which showed the wait itself was the cost rather than the transcription. The wait is now 1.0 second.
• Pasting no longer pauses for nothing. Every paste waited 150 ms for the target app to come forward, even when it was already frontmost and the keystroke was always going to land in the right place. That wait now only happens when focus genuinely has to move.
• Custom replacements now apply on every AI provider. A "replace this with that" mapping (for example k8s to Kubernetes) reached only Apple Intelligence as a hint and did nothing at all on Echo Flow AI, the default on-device model. It is now applied to your transcript before polishing. Terms containing punctuation, such as c++, node.js or .NET, previously never matched at all; they do now.
• Fixed: a polished result could be reused across a reordering. The check that decides whether earlier polish work can be reused compared words as an unordered set, so "Tuesday, not Monday" and "Monday, not Tuesday" looked identical and polish for one sentence could be pasted for the opposite one. Reuse is now refused when the final transcript reorders what the partial already said.
• Includes 1.7.7 (honest update checks). No model re-download.

October 22, 20261.7.7 · build 176

Echo Flow 1.7.7, honest update checks

• Update checks only run where they can succeed. A development build could run an update check and show an update prompt that macOS signature verification would always refuse to install, which looked like automatic updates were broken. The updater is now enabled only in the signed release app.
• Update outcomes are logged. Finding an update, already being current, download failures and aborted updates are recorded, so problems can be diagnosed instead of ignored.
• Includes 1.7.6 (live transcript restored) and 1.7.5 (recording fixes). No model re-download.

October 22, 20261.7.6 · build 175

Echo Flow 1.7.6, the live transcript actually appears

• The live transcript now shows as you speak. The microphone delivers 48 kHz audio, and the resampler used to feed the live decoder finished its conversion after the very first buffer, so every later buffer contributed zero audio. A ten-second dictation accumulated about a tenth of a second — too little to transcribe live — while the file pass still pasted correct text. The resampler is now reset for each buffer; first live text arrives about 2.5 seconds in.
• Includes the 1.7.5 fixes. Switching apps mid-dictation no longer stops the recording.
• No model re-download. Existing installs keep the 229 MB LFM2.5 model.

October 22, 20261.7.5 · build 174

Echo Flow 1.7.5, live transcript restored

• The live transcript is back in the recording panel. With the bundled Echo Flow Speech Engine, the panel could stay on "Listening for your voice…" for the whole utterance even though the pasted text was correct. Two defects in the live-partial decoder caused it: a delay before the first decode that short dictations never reached, and an empty first decode that consumed audio it had not transcribed. A 2.8-second utterance now shows live text in about 1.4 seconds.
• Switching apps mid-dictation no longer stops your recording. The Clinical Note session teardown runs on every app switch, sleep and screen lock, and it was stopping the shared audio engine, hiding the recording panel and forcing the app to idle even with no Clinical Note session running. That teardown is now scoped to a live clinical session.
• No model re-download. Existing installs keep the 229 MB LFM2.5 model.

October 21, 20261.7.4 · build 173

Echo Flow 1.7.4, selected-text rewrite in web fields

• Quick Rewrite now reads selections from web and Electron text fields. Echo Flow reads the system-wide focused element first and asks Chromium-based apps to expose their accessibility tree before falling back to the app-level element. Verified with a 374-character selection in Microsoft Edge.
• A missing Accessibility grant is explained. Instead of silently reporting no text, Quick Rewrite shows an alert with an Open System Settings button. Replacing an app build can invalidate its macOS permission.
• Includes the 1.7.2 cold-start fix. WhisperKit warms at launch and remains the canonical file engine even when streaming falls back to Apple speech.
• No model re-download. Existing installs keep the 229 MB LFM2.5 model.

October 21, 20261.7.2 · build 171

Echo Flow 1.7.2, a rewrite-palette and cold-start fix

• The first dictation after launch is reliable again. The bundled WhisperKit model now warms shortly after launch. If it is still cold when recording starts, live partials use Apple speech but the completed recording is still transcribed by WhisperKit.
• Cold WhisperKit failures fall back cleanly. If WhisperKit cannot run, the file pass uses single-pass SFSpeech instead of producing no text.
• Quick Rewrite explains itself. If there is no recent dictation or selected text, Echo Flow now says so instead of doing nothing. Recent text is intentionally cleared after 15 minutes, a screen lock, sleep, or relaunch; dictate once to seed the palette.

October 21, 20261.7.1 · build 170

Echo Flow 1.7.1, reliability and polish

• Escape works normally again. Escape is no longer captured globally; it cancels only while Echo Flow is recording or showing Preview, so it behaves as usual in every other app.
• One speech engine per session. Echo Flow now chooses WhisperKit or Apple speech before recording starts, avoiding double on-device decoding and mixed live text.
• Apple Intelligence recovers after one bad dictation. A single stalled polish no longer opens the two-minute fallback breaker.
• Recording stops automatically at five minutes. A stuck or forgotten key hold can no longer record indefinitely.
• Safer model downloads and corrupt-model handling. A download is verified before it replaces the working model, and a corrupt local model reports a clear error instead of crashing. Existing 1.7.0 installs keep their model — there is no re-download.
• Cleaner settings. Removed the Paste method picker and the legacy Clear before paste option, and added accessibility labels and tooltips across Settings.

October 10, 20261.7.0 · build 169

Echo Flow 1.7.0, a smaller and sharper on-device model

• A new on-device AI model. Echo Flow AI now runs a fine-tuned LFM2.5-350M cleanup model instead of the previous mumble-cleanup build. On our frozen 432-case internal benchmark, polish improved from 214 to 250 of 288 and rewrite from 48 to 57 of 128, while preserving numbers, identifiers and email addresses noticeably better.
• Smaller and faster. The model download drops from 398 MB to 229 MB, and warm polish is roughly a third quicker (about 66 ms versus 104 ms in the same measurement). It ships as a single pre-fused file.
• Existing installs re-download once. The first launch after updating removes the old cached model and fetches the new one. Until that finishes, dictation falls back to the rules-based cleaner.
• Still entirely on your Mac. No audio or text leaves the device, and no account is required.
• Measured internally, not on your words. A small model can still occasionally alter a word or drop a phrase on unusual input. Preview Before Paste remains the way to check anything important.

October 8, 20261.6.0 · build 168

Echo Flow 1.6.0, privacy controls and Preview reliability

• Choose what ordinary dictation remembers. History saving, spelling-correction memory, vocabulary learning and recent-text recall now have independent controls in Privacy & History. Turning history off offers a choice to keep or delete existing ordinary entries. Clinical Note keeps its separate retention settings.
• Clear recent text explicitly. Ordinary recent text has a fixed 15-minute in-memory lifetime and is invalidated at security boundaries. Clear recent text does not erase saved history, your system clipboard or clipboard-manager copies.
• Recording files follow their actual readers. Temporary audio is removed when recording and transcription consumers no longer need it. Ordinary history no longer stores audio paths.
• Preview keeps the text you reviewed. Edited Preview text is delivered, and authorized working panels survive cleanup and app activation. Clipboard preservation retains supported formats and avoids restoring over newer copies when detected. Paste dispatch is not a guarantee of insertion.
• Your existing settings stay yours. Existing privacy defaults, shortcut, AI model lineup and 1.x license scope are unchanged. Last Dictation recovery is not enabled in this release.

September 15, 20261.5.4 · build 166

Echo Flow 1.5.4, Apple Intelligence polish

• Apple Intelligence is less likely to skip your dictation. Some ordinary transcripts — clinical language, a sharp email, legal or HR wording — were treated as unsafe and silently handed off to Echo Flow AI or the rule-based fallback. Polish and rewrite now use Apple’s profile for editing your own words. Clinical Note classification is unchanged.
• Nothing else about dictation changed. Same shortcut, same overlay, same on-device speech.

September 12, 20261.5.3 · build 165

Echo Flow 1.5.3, rewrite fix

• Correcting a misspelled name no longer counts as losing it. A rewrite that fixed a name the model spelled better — “Mauritus” to “Mauritius” — was judged to have dropped a protected name, so a good result was thrown away and Fix Writing reported “No changes made”. A spelling correction now counts as keeping the name.
• Genuine name changes are still blocked. Dropping a name, or swapping it for a different one (“Mauritus” to “Paris”, “Roberto” to “Roberta”), is still rejected and your original text is used instead.

September 11, 20261.5.2 · build 164

Echo Flow 1.5.2, follow-up fixes

• Tapping the shortcut without speaking no longer looks like a failure. When a recording captured no audio, Echo Flow tried to transcribe an empty file and reported an obscure system error. It now says plainly that nothing was recorded.
• Spoken URLs stay intact. Dictating a URL could glue the following word onto it (https://echoflow.one/docs/for). The address is correct now, and the swallowed word is restored as ordinary text.
• Your words are never replaced by an answer. If a dictated sentence reads like an instruction, the result is no longer pasted — your original transcript is.
• No doubled words across a pause. When the audio file could not be transcribed and Echo Flow fell back to live recognition, a short repeat at the pause boundary could appear twice.

September 10, 20261.5.1 · build 163

Echo Flow 1.5.1, reliability fixes

• Pausing mid-dictation keeps your words. A pause of a second or two could drop the sentence you had just spoken, because the recognizer would start a new segment and discard what came before it. That text is retained now.
• Back-to-back dictations stay separate. Two overlapping processing passes shared a single on-device model state, so a dictation started right after another one — or right after a hotkey warm-up — could come back as the previous dictation, or empty. Processing is strictly serialized now.
• Rewrite actions do what they say. “Make It Shorter” and “Smooth Out” could hand back your text unchanged, or nothing at all. Every action now guarantees a result, falling back to Echo Flow’s own formatter when the AI output can’t be used. Bullet points come back as bullets.
• Cleaner polishing, every time. Fillers, false starts, and immediate self-corrections are removed, names, numbers, and URLs are preserved, and polishing is deterministic: the same words produce the same result.

August 22, 20261.5.0 · build 160

Echo Flow 1.5.0, Liquid Glass & HIG pass

• Native Liquid Glass on macOS 26. The main window and all overlays (recording HUD, rewrite palette, preview) now use native macOS 26 Liquid Glass, with an adaptive vibrancy fallback on older macOS versions.
• A real Mac menu bar. Full App / Edit / View / Window / Help menus with standard shortcuts (⌘H hide, ⌃⌘F full screen, ⌘? help).
• Resizable windows. The main window is no longer locked to a fixed size, resize freely; History dropped its always-on-top level.
• Native system colors. Text, separators, and status indicators follow your appearance and Increased Contrast settings.

August 19, 20261.4.1 · build 158

Name-preservation guard

• AI polish can no longer silently swap your names. If a polished output drops a proper noun, number, URL, or email from your dictation, Echo Flow rejects it and pastes the raw transcript instead.
• Reliability hardening across the menu bar, history search, and accessibility paste path.

August 17, 20261.4.0 · build 157

Live transcription for every dictation

• See words while you speak, in all sessions. WhisperKit dictations now stream live partial text into the recording overlay as you talk, matching what Apple Speech sessions already did.
• Hallucination guard. Decoder-confidence risk scoring forces preview on error-prone decodes before anything reaches your text field.
• Faster paste after release. Speculative polish now runs on the final transcript itself, so polished text lands near-instantly after you stop talking (v1.3.4).
• Recovery from mid-recording audio device changes, plus an Apple Intelligence circuit breaker that falls back instantly if the system model stalls.

August 12, 20261.3.4–1.3.5 · builds 155–156

Polish quality hotfixes

• Good AI polish is no longer thrown away. Fixed an overly strict sanity check that rejected legitimate grammar fixes and pasted the raw transcript instead.
• Faster, steadier polish: fresh per-request Apple Intelligence sessions and tighter polish deadlines cut wait time and eliminated context-window failures.

August 7, 20261.3.3 · build 154

Searchable history

• Full-text search across your archive. Word and prefix search with relevance ranking across every saved dictation, with graceful fallbacks for short queries.

August 1, 20261.2.0 · build 147

Echo Flow 1.2.0

• Long dictations no longer drop their opening. File transcription now scales its timeout to the recording length, so 40–60 second (and longer) dictations finish transcribing instead of falling back.
• Versioning moves to 1.2.0 for clearer release management after the Clinical Note launch.

July 31, 20261.1.0 · builds 145–146

Clinical Note mode (+ hotfix)

• Dictate a 2–10 minute summary, get a structured note. Press ⌥⌘N, speak a post-session summary, and Echo Flow organizes it into a SOAP, DAP, BIRP, GIRP, Progress, or Intake note.
• Review before it lands. Every note opens in a full editor, edit, paste into the field you were using, copy, or discard. Clinical notes never auto-paste.
• Patient data stays on your Mac. Organization runs on-device, recordings are deleted as soon as they're transcribed, and only the raw transcript (never the generated note) is stored locally.
• Built for the field you already use. Works with any web-based practice-management system, no integration or API needed.
• Made for long dictation. A dedicated toggle shortcut, an 8-minute warning, and a hard 10-minute cap keep long recordings reliable.
• Easy to turn off: the Clinical Note toggle lives in Settings, and retention of raw transcripts is independently configurable (Off, 7, 30, 90, or 180 days).
• Cancel now closes the recording overlay in every terminal path: cancel, cancel during transcription, start failure, and app quit.

July 28, 20261.1.0 · build 144

Echo Flow 1.1.0

• Reliability and direct-distribution improvements.

June 24, 20261.1.0 · build 130

Smoother rewrite palette

• Smooth Out replaces Make It Longer, joins fragments into full, flowing sentences without inventing new content.
• Sound Pro replaces Sound Professional, shorter tile label, same clear, direct, work-ready tone.
• Every rewrite action got a fresh, action-specific prompt. Fix Writing, Make It Shorter, Sound Friendly, Sound Confident, Bullet Points, Key Takeaways, Keep Original, and Plain English all re-tuned.
• Enum case names and saved-action IDs are unchanged, so your existing Settings, hidden actions, and palette layout carry over.

June 2026

Smarter polish, clearer history badges

• Self-correction repair. When you correct a word mid-sentence ("the returns, the results we are getting"), Echo Flow AI now keeps only the corrected form.
• New "Polished (no changes)" badge. History now distinguishes "polishing ran but had nothing to fix" from "polishing was skipped".
• New Echo Flow Speech Engine. Replaced the bundled Moonshine engine with our own on-device speech engine, tuned for dictation and running on Apple Neural Engine.
• Visual consistency sweep. Settings, Updates, Snippets, and Rewrite Modes now share the same color, font, spacing, and corner-radius tokens.

June 2026

Echo Flow AI is now one focused model

• One purpose-built local model replaces the previous Lite/Pro lineup. Fits any Apple Silicon Mac with 4 GB RAM, including base M1 and M2 with 8 GB unified memory.
• Fine-tuned on real dictation. Trained on tens of thousands of voice-dictation transcript pairs, not chat data.
• Runs entirely on your Mac after a ~380 MB one-time setup. Your dictated audio, transcripts, and rewritten text stay on your Mac.
• Fast first-token response and a clean, quiet install. The same local model handles everyday polish and every rewrite action.

June 2026

Echo Flow AI is now available

• Echo Flow AI now powers local polish and rewrite after a one-time setup download.
• Rewrite selected text in any Mac app with the 9 practical rewrite actions.
• Your dictated audio, transcripts, and rewritten text stay on your Mac.
• History now tracks which polish provider handled each entry.

May 2026

Launch-ready Mac dictation workflow

• Hold Option-Space to dictate into any Mac app.
• On-device speech recognition with local polishing fallback.
• Smart Context adapts tone for mail, chat, notes, code, and more.
• Preview Before Paste lets you paste, edit, or discard output.

May 2026

Rewrite Palette and Your Styles

• Rewrite selected text with Clean Up, Make It Clear, Tighten, Turn Into Email, Slack Reply, Action Items, Quick Recap, Keep Verbatim, and Plain English.
• Custom user styles appear directly in the rewrite palette; palette buttons can be shown or hidden in Settings.
• If replacement fails, Echo Flow copies the result to your clipboard.

May 2026

Snippets, history, and personal dictionary

• Voice-triggered snippets support date, time, and clipboard placeholders.
• Searchable history supports open, copy, add term, and refine again actions.
• Personal Dictionary protects names, acronyms, products, and exact phrases.
• Latency improvements reduce the time from release to pasted text.