Skip to content
Start Free Trial

Release notes

Every release of jhint, newest first — macOS and Windows in one list, each note marked with the platform it belongs to. The app checks the same list when it looks for updates.

Version 1.8.6 Windows

8 September 2026

  • A model that cannot answer is no longer accepted silently. When you pick a model — from the list or by typing its id — jhint asks the provider whether it answers, and keeps your previous model if it does not, saying why in the provider's own words. Before this you found out during a conversation, on the first answer.

Version 1.8.5 Windows

8 September 2026

  • The newest OpenAI models answer again. They reject the field jhint used to limit the length of an answer and ask for a different one; jhint now takes the name from the service's own reply and repeats the request with it. Before this, picking such a model produced an error instead of an answer.
  • The model list only offers models that can answer. A provider lists everything it has — embeddings, speech synthesis, transcription, image models — and jhint offered all of them, although none can hold a conversation. You can still type any model id by hand.

Version 1.7.4 macOS

8 September 2026

  • Answer length is now capped by the name the provider itself asks for. Some newer models reject the usual name and say which one they want; jhint repeats the request with it instead of showing you an error. This also covers your own server behind “OpenAI-compatible”, where the provider name tells us nothing.
  • The model list drops everything that cannot answer, wherever the giveaway word sits in the name — speech, transcription, embeddings, image and moderation models no longer appear as choices.

Version 1.7.3 macOS

8 September 2026

  • Checking a model on ChatGPT API works again. The check was sending a parameter the newer models reject, so a perfectly good model came back as unavailable and the selection fell back to the previous one.
  • The model list now shows only models that can answer. Providers list everything they have — embeddings, speech recognition, image models — and those could end up selected as your answering model.

Version 1.8.4 Windows

1 September 2026

  • A licence check now also says which platform the app runs on — one word, sent with the check the app already makes. It exists so we can see how installations divide between macOS and Windows; nothing else about your computer is sent: no Windows version, no hardware. The Privacy Policy lists every field a licence check carries.
  • Three more open-source components are named inside the app. ColorCode, NuGet.Versioning and the Windows SDK .NET projections travel with jhint and were missing from Settings → General → Diagnostics → Open Source Licences. Nothing about the program itself changed.
  • If you already have 1.8.4 installed, updates will keep working as usual. Copies of 1.8.3 are the exception: the key their update check trusts was replaced, so they will not accept this update and need to be installed again from the download page. Your licence and your saved data stay where they are.

Version 1.7.2 macOS

1 September 2026

  • A licence check now also says which platform the app runs on — one word, nothing about your computer. It exists so we can see how installations divide between macOS and Windows; the Privacy Policy lists every field a licence check carries.

Version 1.8.3 Windows

31 August 2026

  • First Windows release. jhint listens to the call on your computer, recognises what is said and answers — the same product as on macOS. One licence key works on both.
  • Answers on your own machine. The built-in model needs a discrete graphics card: measured 79 tokens/s on an RTX 4060 against 14 on the processor and 12 on integrated graphics, with 15 as the threshold for keeping up with a conversation. Without a discrete card the built-in model is not offered — use an external provider, or Ollama or LM Studio on your own machine.
  • Speech is recognised in the language you speak, detected on its own — there is nothing to choose. Checked on English, German, French, Spanish, Russian and Italian.
  • The overlay sits over the call. It is see-through as far as you set it, resizes from any edge or corner, and while the model writes it holds the start of the answer at the top — until you scroll, and then the window is yours.
  • Nothing about your conversations is stored until you turn it on. When you do, they are encrypted with a key kept in Windows Credential Manager, and you choose how long they live.
  • Uninstalling removes everything — downloaded models, saved conversations, provider keys and this computer's place in your licence — whether you start it from the settings or from Installed apps.
  • The installer is not code-signed yet. Windows will warn you the first time you run it; you can choose to run it anyway. A certificate is planned once there is demand for it.

Version 1.7.1 macOS

31 August 2026

  • The answer window is yours the moment you touch it. While an answer is being written jhint keeps its first line at the top of the window; scroll yourself and it stops following until the next answer.
  • The overlay is now always dark, whatever theme the system is set to. On a light theme its text used to be barely readable through the glass.

Version 1.7.0 macOS

29 August 2026

  • Speech recognition rebuilt: every word now belongs to exactly one reply, so replies no longer repeat or swallow each other.
  • The end of a reply is no longer echoed as a second line.
  • A guard against invented text now also covers the middle of a reply, not only its tail.
  • Pressing Answer sends exactly what you heard up to that moment — speech recognised while the tail is being ground goes into the next answer.
  • The assistant’s behaviour now lives in the System prompt field, where you can see and change it; only safety rules stay hidden.
  • When a provider cannot be used, the exact reason is shown instead of a generic message.

Version 1.6.8 macOS

28 August 2026

  • The first answer is no longer the slow one. Your session instruction and the files you attached now go into the built-in model while the session is starting, so the model has already read them by the time you ask. On a session with a page of instructions that took about two and a half seconds off the first answer; every answer after it was always fast, and now the first one matches.
  • Sessions start sooner. Speech recognition and the built-in model used to be prepared one after the other. They are prepared side by side now — the wait is as long as the slower of the two instead of both added up — and the window says which of them it is still waiting for.
  • Long answers stay readable. A new answer used to appear at the bottom of the window and its first lines slid out of sight as the text grew. Now the answer rises until its beginning meets the top edge, and only then does the text continue downwards.
  • Models that will not fit are no longer offered. Each answer model says how much memory it needs, and that is now checked against the memory your computer actually has — together with the model that transcribes the call, since both are running at once. A model that would not fit is shown with the numbers and the reason instead of downloading five gigabytes to find out.

Version 1.6.7 macOS

27 August 2026

  • Claude Code and Codex are gone from the provider list. They worked through the command-line tool you had installed, and that path was slow: about three to eleven seconds before the first word, against roughly one second straight from an API. In a live conversation that wait is the whole problem, so the two options have been removed rather than left as a trap. If you were using one of them, pick the model that ships with jhint, or an API key with the same vendor — Claude API and ChatGPT API answer far faster.
  • Screenshots never touch the disk. Sending one through a command-line provider used to require writing the image to a private temporary folder first. That path no longer exists: a screenshot now lives only in the session’s memory and inside the request itself.

Version 1.6.6 macOS

25 August 2026

  • Answers now appear as they are written. The assistant used to hold the whole reply until it was finished — on a long answer from the built-in model that meant watching an empty box for about fifteen seconds, even though the first words were ready in a fraction of a second.
  • Starting a session right after stopping one no longer breaks speech. If you pressed Answer and then restarted the session, the new one could run with no transcription at all until you restarted a second time.
  • Technical terms keep the spelling you gave them. Names from your session prompt, context files and profile — GStreamer, PostgreSQL, H.264 — are now written the way you wrote them, instead of “G-Streamer” or “GST Web App-C”.
  • On the Medium model the transcript no longer freezes. Text used to stall for seconds at a time while the full phrase was being recognised; drafts are now handled by a lighter model, and the reply you send is still the accurate one.

Version 1.6.5 macOS

24 August 2026

  • The provider list no longer contradicts itself on Individual. Providers that need Pro are shown but greyed out, so picking one can’t leave you with a heading about Claude while Embedded stays selected. If Pro would add something for you, one line and an Upgrade button now sit right under the list.

Version 1.6.4 macOS

24 August 2026

  • The built-in models answer far faster. Qwen3 8B replied to a short line in 12 seconds; it now takes under a second — it was spending that time on internal reasoning you never saw. Long answers are three to seven times quicker across all three models.
  • No more repeated sentences. On longer answers Qwen3 VL 4B could repeat the same phrase until it hit the length limit, taking 45 seconds. That’s fixed: the same question now answers in 8 seconds, once.
  • Screenshots are read twice as fast by the built-in model — 2.6 seconds instead of 5.9. Answers from providers you bring yourself are unaffected: they keep the larger image, where small text matters more than seconds.

Version 1.6.3 macOS

24 August 2026

  • Models with two files download again. Qwen3 VL reads screenshots and ships as weights plus a projector. If the weights finished and the projector did not, resuming asked the server to continue past the end of a complete file and failed with “HTTP 416” — the model could not be downloaded at all. A finished file is now verified by its checksum instead of being re-requested, and a refused resume quietly starts that file over.

Version 1.6.2 macOS

24 August 2026

  • Both windows now tell the same story. A download started in Settings shows the same progress in the overlay, and the same percentage — no more “Paused at 79 %” next to a live download at 62 %, and no more “Choose STT model” while that model is downloading.
  • No surprise wait when you decide to talk. A speech model that is downloaded but not yet prepared no longer offers Start session; it offers Prepare, and says the first run takes a few minutes. The overlay button downloads the model right there instead of sending you to Settings.
  • Your choice of answer model stays chosen. Picking a model while another one is downloading, or changing a setting mid-download, no longer reverts the choice or cuts the download short. A partly downloaded model offers Continue with the size left, not the full download again.
  • Downloads are always visible. Whatever else the overlay is asking for, running downloads are listed underneath with progress and a Cancel button. Picking a model works by clicking anywhere on its row.

Version 1.6.1 macOS

23 August 2026

  • Setup asks for your licence first. The wizard now goes permission → licence → engine, and nothing is downloaded until the licence is active — no more waiting on a 2.5 GB model you cannot use yet.
  • The setup window stays until setup is done. It no longer closes halfway into a half-configured app; quit any time with Cmd+Q and jhint reopens on the very step you left. Reach the final screen and setup is complete.
  • The screen recording prompt appears once. On a first run it used to come back two or three times in a row — each internal check re-asked macOS. Now you see it when you ask for it, and not again.

Version 1.6.0 macOS

22 August 2026

  • jhint now comes in two plans. Individual runs the built-in model on your Mac — nothing to sign up for. Pro adds everything else: Ollama, LM Studio, your own API keys with Anthropic, OpenAI, Google, Groq or OpenRouter, and any OpenAI-compatible endpoint. Existing licences continue as Pro.
  • Change your plan from the app — Settings → License → Change plan. You see what is charged today and what your plan costs from then on, before you confirm. Moving down keeps your keys and settings on the Mac; they start working again if you move back.
  • The Provider tab shows every provider, marking the ones your plan doesn’t include — so you can see what you would be paying for, instead of a list that quietly grows after a purchase.
  • Individual covers one Mac, Pro covers two. If a Mac is gone and its place is still taken, «Use this Mac instead» frees it — the machine that has been offline longest gives it up.
  • Enterprise is now a conversation rather than a checkout: jhint.ai/enterprise.

Version 1.5.0 macOS

21 August 2026

  • jhint now answers on its own. A model runs on your Mac: no account, no API key, nothing to sign up for. Choose it on first run, wait for the one-time download, and you are working. Your own provider is still there if you prefer one.
  • A setup window on first launch walks you through it — one permission, how jhint should answer, your licence key — while the model downloads in the background.
  • The built-in model can look things up on the web when you allow it. Searches run from your Mac, over your own connection, and every query it makes is shown to you. Off until you turn it on.
  • Settings open instantly — the window is now prepared in advance instead of being built on the click.
  • The AI server no longer outlives the app: whatever happens to jhint — a crash, Force Quit, logging out — the model is released from memory with it.

Version 1.4.2 macOS

18 August 2026

  • jhint now tells you when the local AI server isn't installed. Ollama is selected out of the box, and if it has never answered, jhint says so plainly and offers a button that takes you to where you get it — instead of a bare "not responding" and a Try again that had nothing to try. If the server has worked before and is simply not running, nothing changes: you still get Try again.
  • Settings link straight to the download page for Ollama and LM Studio, next to the server address — the same place external providers have offered "Get an API key" all along.

Version 1.4.1 macOS

18 August 2026

  • Nonsense text no longer reaches the screen. On unclear audio the speech model could get stuck repeating one word or one letter until it ran out of room — dozens of copies in a row, sent on to your AI as if you had asked for them. Such runs are now caught and dropped, and the run is cut short the moment it starts looping instead of stalling the line for seconds.
  • Dutch sessions now get the foreign-script guard that every other alphabet language already had.

Version 1.4.0 macOS

18 August 2026

  • jhint can now uninstall itself. Settings → Privacy & data → Uninstall jhint removes the speech models, settings, saved conversations, provider API keys and your licence key from the keychain, frees this Mac's licence slot so you can use the key elsewhere, and moves the app to the Trash. Screen Recording and Microphone permissions stay in System Settings — macOS keeps those for apps it no longer sees, and the dialog says so.
  • Check for updates without hunting through the menu. The version line in Settings → General has its own button now. Worth knowing: automatic updates install when the app quits, and an app that lives in the menu bar rarely does.

Version 1.3.1 macOS

18 August 2026

  • A finished phrase now appears the moment the speaker stops. Until now it waited for the full-phrase pass — about a second and a half — and then landed as a block. It shows up straight away and is quietly replaced with the precise text when the pass finishes.
  • Recognition no longer stalls mid-speech. On an unclear passage the decoder could get stuck repeating itself and hold the screen for seconds while you kept talking; such a run is now cut short. The text it produces is unchanged.

Version 1.3.0 macOS

18 August 2026

  • jhint has moved to jhint.ai. Updates, licence activation and the support address now live on the new domain. Nothing to do on your side: this update switches the app over, and the previous addresses keep working for earlier versions.

Version 1.2.5 macOS

17 August 2026

  • Spreadsheets and presentations can be attached to a session. Excel and PowerPoint files are read on this Mac: cell values with their dates, times and percentages, every sheet in the workbook, the text of each slide and the speaker notes. Charts and pictures inside them are not read — that needs Excel or PowerPoint itself.
  • Attached RTF and HTML files now reach the model as documents. Until now their markup went into the prompt instead of their text, tags and stylesheets included.
  • A very large attachment no longer breaks every question. Only the first 200,000 characters of a file are used, and the dialog says which file was cut. Before, an oversized file made the provider reject each request in the session with no explanation.

Version 1.2.4 macOS

16 August 2026

  • Granting Screen Recording really does one thing per click now. The first click asks macOS for access and nothing else; System Settings opens only on the next click, after you have refused. Until now jhint could still open Settings on top of the system alert, leaving Settings running with no window until you logged out and back in — the fix shipped in 1.2.3 never took effect.

Version 1.2.3 macOS

16 August 2026

  • Set jhint up while the speech model downloads. The first download takes minutes. Whatever still needs doing — a licence, Screen Recording, a provider key, a model — is now shown during the download instead of after it, and each button opens the exact place it is fixed.
  • Granting Screen Recording no longer leaves System Settings blank. jhint asked macOS for access and opened System Settings at the same moment, which could leave Settings running with no window at all until you logged out and back in.
  • Settings no longer call the speech model downloaded while it is still downloading. Speech recognition showed “downloaded” and offered to prepare a model that was half-way through its download; it now reports the download that is actually running.

Version 1.2.2 macOS

15 August 2026

  • Open source licences are now in the app. The full licence text of every open-source component jhint is built on is included in the app and readable in Settings → General → Open Source Licences.
  • A word before you save a conversation. Saving conversations can require the knowledge or consent of everyone involved, and the rules differ from place to place. jhint now says so before you turn saving on, and once before your first session.

Version 1.2.1 macOS

15 August 2026

  • Offline access follows your paid period. jhint keeps working without a connection until the end of the period you have paid for, and for up to seven days at a time.

Version 1.2.0 macOS

15 August 2026

  • Quiet speakers are heard. Speech is now measured against the background of your call instead of a fixed loudness threshold. A softly spoken colleague, a distant microphone or a quiet conference line used to be treated as silence and skipped entirely; on a five-minute recording that meant two thirds of the speech never reached recognition at all.
  • Nothing is dropped while the recogniser is busy. Audio keeps flowing in while a phrase is being transcribed, so speech that lands during recognition is no longer lost.
  • Sentences are no longer cut mid-thought. A pause in the middle of an unfinished question — “when did you last… look at the numbers?” — keeps the question together. jhint waits when the sentence sounds unfinished and closes it when it sounds complete, in any language.
  • Answer sends everything you heard. Pressing Answer now finishes recognising whatever was said up to that moment, so the last second of a question no longer gets left behind.
  • New History tab. Conversations you chose to keep are listed with their date, length, number of questions and the model that answered. Open one to read it, export it to Markdown, or delete it.
  • Stored conversations are encrypted. They are kept on your Mac, encrypted with a key held in your keychain, and nothing is written at all when “Store transcript” is off or “Keep data for” is set to “Do not keep”. “Keep data for” now really removes conversations once their time is up.
  • Clear context. When the conversation drifts to something irrelevant, one button — or ⌃⌥C — drops what has been said so far so it is not sent with your next question. Your stored conversation stays intact.
  • Windows remember where you left them. Settings and New session open centred the first time, then reopen at their previous position and size.
  • Various bug fixes.

Version 1.1.6 macOS

14 August 2026

  • The model list now shows only models you can actually talk to — speech-to-text, text-to-speech and classifier models are filtered out, and each remaining model says whether it takes text or text and images.
  • Screen capture no longer fails silently: a model that cannot see images says so instead of quietly dropping the screenshot.
  • Starting a session checks that the selected model is answering, and explains the problem before the conversation begins rather than during it.
  • The overlay shows which provider and model will answer, both before and during a session.
  • Picking a provider now always leaves a model selected instead of an empty box.
  • Changes made in Settings reach the overlay immediately, without closing the window.
  • More room for the answer: the recognised speech area is smaller.

Version 1.1.5 macOS

14 August 2026

  • Activating a licence key now updates the overlay right away, instead of leaving the "Enter licence key" button until Settings was closed.

Version 1.1.4 macOS

13 August 2026

  • Campaign supporters now see their badge on the License tab.
  • The end of a fixed-term licence shows the year and how long is left.

Version 1.1.3 macOS

12 August 2026

  • The app now talks to the licence service over its own domain. Nothing changes for you — activation, renewal and key recovery work exactly as before.

Version 1.1.2 macOS

12 August 2026

  • Lost your licence key? Settings → License now has a "Lost your key?" form: enter the email address you paid with and the key is sent to it again. Until now the only way was to write to support and wait for a reply.
  • Settings → License shows the email address your licence is registered to, so you know where that key will arrive.

Version 1.1.1 macOS

12 August 2026

  • Granting Screen Recording is one button now: it asks the system for access and opens System Settings on the right row. Before the first request macOS does not list the app at all, so the settings page had nothing to switch on.

Version 1.1 macOS

12 August 2026

  • Web search: turn it on in Settings and the model looks things up when a question needs current information. Answers that used it show a “searched the web” mark.
  • Answers are formatted: headings, lists, tables and code instead of raw markdown.
  • New app icon and menu bar symbol.

Version 1.0.3 macOS

8 August 2026

  • New app icon and menu bar symbol.
  • The menu now shows the product name and version.

Version 1.0.2 macOS

8 August 2026

  • jhint now tells you what is missing before a session — permission, API key or model — instead of failing once the conversation has started.
  • A saved API key is shown as a masked value, so it no longer looks like it was lost.
  • Screen Recording shows one button that actually works, and the status updates when you come back from System Settings.
  • The Settings window no longer disappears when you switch to another app.
  • Only one copy of jhint can run at a time.

Version 1.0.1 macOS

7 August 2026

  • Verification build for the update mechanism. No functional changes.

Version 1.0 macOS

7 August 2026

  • First public release.

Updates install from within the app. The latest build is also on the download page.