macOS
Every Scribe0 for macOS release, in plain language.
Download .dmgv0.3.13
Added- Literature review from dictation. Ask for a literature review while dictating ("look up the evidence for...") and a chip appears carrying the research question. Tap it to search Consensus (connect your account in Settings > Integrations), browse the papers, add citations to the Context Cart, or insert a references block into the note. Only the question shown on the chip ever leaves your Mac, and only when you tap it.
- SmartEdits (Pro). Scribe0 learns corrections you make repeatedly after enhance and applies them automatically from then on. Every applied edit shows a receipt chip that can undo the change or turn the rule off.
- Microphone picker. Choose which input device Scribe0 records from in Settings > Microphone.
- Claude Opus 4.6 is available in the Bedrock model picker.
Changed- The enhance status line always names the backend drafting the note.
- On-device model loads now say what they are for (for example "Action detection"), so a local model warming up beside a cloud enhance is not mistaken for the whole pipeline running locally.
- The Settings window is wide enough for all its tabs instead of collapsing the last few into an overflow menu.
- The MediPhi on-device model has been retired from the model picker; its download source is no longer available. Qwen3 4B remains the on-device model, and any previous MediPhi selection falls back to it automatically.
Fixed- Suggested-action chips (referrals, follow-ups, literature review) now appear for transcripts pasted in with hard line breaks. Detection previously ran but every suggestion was silently dropped.
- SmartEdits failures are surfaced instead of skipped silently, and the missing rewrite rule type was added.
v0.3.12
Added- A command bar for editing notes by instruction. A new bar under the workbench lets you tell Scribe0 how to change the note you are looking at. Type or dictate an instruction such as "make it more concise" or "remove the medication list", and Scribe0 rewrites whichever tab is open, adding the result as a new tab so the original is preserved. Dictation uses the same on-device speech recognition as your notes. Editing by command is a Scribe0 Pro feature.
v0.3.11
Changed- Dictations now save to history automatically. There is no need to press Save: a finished transcription, an imported file, an enhanced note, and any generated letter or follow-up list are all kept as you go. Re-enhancing a note, or restoring an earlier one and enhancing it again, updates that same history entry instead of creating a duplicate.
- Closing the window now quits the app. This releases the on-device AI model's memory (about 2.3 GB held on the graphics processor), which previously stayed reserved until you quit. Your work is safe because dictations are saved to history automatically.
- Enhanced notes no longer contain em-dashes. Long dashes are replaced with a spaced hyphen, so notes read consistently whichever AI backend wrote them.
Fixed- Clearing the workbench now also removes a dropped referral and any attached context, even if the app was still reading them. Previously one could reappear a moment after you cleared.
- Starting a new recording no longer discards a referral you just dropped in. Recording continues the current note, so you can dictate a reply with the referral still attached.
v0.3.10
Added- A medication safety net on the on-device enhance path. Speech recognition often garbles drug names, and a small on-device model can quietly write a plausible wrong one. Scribe0 now matches garbled names against an Australian medication lexicon and offers them to the model as spelling corrections, flags any medicine that was invented or dropped once the note is written, and highlights names it could not resolve so you can click one and pick the drug you meant.
- A "Clinical polish (local)" prompt tuned for the on-device model, so it reliably returns the full polished note as well as the plan.
- Rendered markdown for enhanced notes, with a button in the note toolbar to switch between the formatted note and the raw text.
- The enhanced note now streams in as the model writes it, on the on-device, Apple Intelligence and Bedrock backends.
Changed- The fidelity guard now tells the model never to guess at a garbled medication name, and to keep the dictated wording marked unclear instead.
- A Pro licence now identifies its device to the enhance proxy, which binds each licence to a small number of devices.
Fixed- Confirming a detected referral recipient now writes the addressed letter immediately, instead of waiting for another Enhance.
- The medication spelling guidance no longer leaks into the finished note.
- The everyday word "cleanse" is no longer flagged as a medicine.
v0.3.9
Security- Saved dictation history is now encrypted on disk. This matters because saved documents can now hold patient names and dates of birth. The encryption key lives in the system keychain.
Added- Drop a referral letter into the Context Cart. A PDF referral, scanned or digital, is read entirely on your Mac: Scribe0 uses the PDF's own text layer where it has one and falls back to on-device Apple Vision OCR for scanned pages. The referral then seeds the note's context with no retyping.
- Scribe0 writes the letter back to the referrer. When you drop a referral, it works out who referred the patient and offers to address the reply to them. Confirm the chip and your enhanced note becomes a properly addressed letter: a salutation, the recipient block, an RE: subject line, a sign-off, and your letterhead. The referrer's name, the patient's identifiers, and the CC list are added on your Mac AFTER the AI step, so third-party names and patient details are never sent to the cloud.
- Your clinician profile. Set your name, credentials, specialty, and practice once in Settings, Profile. Letters are then signed and framed with your letterhead, and the profile is what lets Scribe0 tell the referrer apart from you.
- Patient details carried onto the reply. The patient's name, date of birth, and the CC list are pulled from the referral and placed on the letter as an RE: line and a cc: line, each checked against the source document so nothing is invented.
- Run the AI entirely on your Mac. A new Local LLM switch in the main panel routes the enhance step to an on-device model (Qwen3-4B), so a note never leaves your machine. Your choice is remembered between launches.
- Repeat visits are grouped. When a document identifies a patient, it is saved with that patient and grouped with their earlier visits in History, with a visit count.
Changed- The on-device language model download now shows a percentage bar with a Stop button, so a multi-gigabyte download reads as progressing rather than hung.
Fixed- The on-device language model (about 2.3 GB) could time out and fail part way through downloading on the first run. Large model downloads now run to completion.
v0.3.8
Added- You can stop things now. A model download, an enhance, a referral letter being written, or the search for referrals and follow-ups: each can be stopped while it runs. Enhance turns into a Stop button while it is working, the download banner has a Stop, and tapping a chip that is mid-letter stops it. Stopping is not treated as an error, and never leaves a half-written note in a tab or a half-downloaded model on disk.
Fixed- The on-device model download looked stuck at 0%. It was not stuck: the bytes were arriving the whole time, but the progress readout never moved, so a download that takes several minutes (and much longer on a slow connection) was impossible to tell apart from a hang. The readout now moves as the model downloads and shows the total size, so you can see it working and judge how long it will take.
- If the model download does fail, Scribe0 now says so instead of quietly retrying against a slower source that could hang indefinitely with nothing on screen.
v0.3.7
Added- Scribe0 now spots the referrals in your dictation and writes the letters. Say "I'll refer her to endocrinology" during a consult and, once the note is enhanced, a suggestion appears under the note tabs. Tap it and the referral letter lands in its own tab, addressed "Dear Doctor", drawing on the note and the dictation. Say "chase her thyroid function next week" and you get a follow-up checklist the same way.
- Detection understands the difference between a referral you are making and one you are only talking about. "I referred her last year", "the GP already referred him", and "if her eGFR drops I'll refer" all produce no suggestion. Nothing is ever written without your tap.
- Detection runs on device, for everyone, using the same Qwen3 model as the on-device enhance. It costs nothing and no transcript leaves your Mac to produce a suggestion. Writing the letter itself is a Pro feature.
- Settings, Actions. Edit what Scribe0 looks for and how it writes each document. Scribe0 always applies its own fidelity rules on top of your wording: it will never alter a dictated dose, date, or result, and never invent a recipient's name.
v0.3.6
Security- Dictated numbers, doses and dates are now protected from being "corrected". Small on-device models will silently rewrite a dictated value to the one they saw most often in training. A dictated "aspirin 300 mg" (the AU and UK loading dose) came back as "325 mg" (the US one), and in one case the model rewrote a dictated calcitriol dose and labelled the original as a typo. Every enhance now carries fidelity rules that forbid this, applied by the app itself rather than by the prompt text. That means the protection also covers prompts you write yourself, not only the ones that ship with Scribe0. Nothing to reset, no action needed. This reduces a measured failure mode. It does not make an on-device model safe to trust unread, so always check clinical values against what you dictated.
Added- Qwen3-4B-Instruct is the new on-device AI model, replacing MediPhi as the default for the on-device (MLX) enhance backend. It produces cleaner, better structured notes, follows formatting instructions far more reliably, and finishes about 2.4x faster. MediPhi is still selectable under Settings, Enhancement Backend.
- A time estimate while enhancing. The progress bar now shows roughly how long an enhance will take, learned from your own past runs. It stays an indeterminate spinner until it has enough samples to give an honest answer rather than a made up one.
v0.3.5
Fixed- The context cart popover was hard to read in dark mode.
- The app version is now derived from a single source, so About and the updater can no longer disagree about which version you are running.
v0.3.4
Added- On-device AI enhance (MLX). A medical language model (MediPhi 3.8B) that runs entirely on your Mac's GPU. No network, no cloud, no cost. The model downloads once from Scribe0's own mirror on first use.
- Prompt gallery. Browse, search, import, update and remove prompts.
- Per-prompt backend and model. A prompt can pin the backend and model it runs on, so a proofread prompt can stay on-device while a summary goes to the cloud.
- The enhance status now shows a phase-aware animated icon that distinguishes sending from receiving.
v0.3.3
Added- Automatic updates, via Sparkle, with a "Check for Updates" menu item.
Fixed- The transcript and history panes now fill the window when it is resized.
v0.3.2
Changed- The speech model now downloads from Scribe0's own mirror instead of HuggingFace, with a progress readout on first run. HuggingFace's storage backend was stalling the download part way through, so first launch could appear to hang.
v0.3.1
Fixed- Pro licence keys were rejected with a 401 in some cases. Keys are now validated when saved.
- The microphone permission prompt never appeared on some fresh Macs, leaving dictation silently unable to record.
v0.3.0
Added- Free and Pro plans for the cloud enhance step.
- Sovereign AWS Bedrock (Australia) is now the default enhance backend when it is configured, keeping transcripts in Australian regions.
Fixed- Long enhancements no longer fail with a "connection lost" error. The Bedrock response is now streamed.
v0.2.0
Added- Import an audio file (
.m4a,.wav,.mp3and others) and transcribe it offline. - Sovereign AWS Bedrock (Australia) as an enhance backend.
- Speaker labels are now attributed at word granularity.
- On a shared Mac, the speech model and API key can be shared across user accounts.
Fixed- An expired or invalid OpenRouter key now warns instead of failing silently.
- Import an audio file (
v0.1.0
First public release. A notarized, Apple Silicon native dictation app.
Added- Local speech to text on the Apple Neural Engine (Parakeet TDT v2). Audio never leaves your Mac.
- AI Enhance with a choice of backends: Apple's on-device Foundation Model, or the cloud via OpenRouter.
- On-device PII redaction and drug name correction, run before any cloud call, with the detected spans highlighted in the original transcript.
- Speaker diarization, editable transcripts, prompt presets, a context cart, text replacements, and session history.
- Per-stage timings and diagnostics.