Jauvex Personal 1.8.3

Jauvex Personal 1.8.3 for macOS (Apple silicon): open jauvex-personal-1.8.3-mac-arm64.dmg, drag Jauvex to Applications and open it. The app is signed and notarized by Apple.

On macOS or Linux you can also install it with one command: curl -fsSL https://jauvex.ai/install | sh. On Windows, in PowerShell: irm https://jauvex.ai/install/windows | iex.

SHA-256 of the disk image: af0c24ee2314fb2762adf70d596da785de8cf26975372097ef921777040a2bdc

What's new

• Safer under the hood: the app moves to Electron 45.2.3 (Chromium 156.0.8078.80, Node.js 24.22.2), with the browser engine's
  security fixes of late October, among them flaws that let a crafted web page run code outside the browser's sandbox.
• Claude Fable 5.2: Anthropic's newest model can be picked for your Claude agents.
• Sonnet 5.5 and Haiku 5.5 can be picked for your Claude agents too.
• Codex 0.167: Codex agents signed in with an OpenAI API key can pick GPT-6.1 Terra, a compaction that fails no longer leaves part
  of a summary in the conversation, and Codex's commands run in a sturdier sandbox.
• Fixed: a message you send just as a Claude agent finishes its answer now gets an answer of its own, instead of looking ignored.
• Fixed: a simple yes-or-no question from an agent shows Yes and No buttons again.
• Fixed: when you dictate to a board's or a workflow's chat, your message no longer shows "[voice transcript]" in front of your words.
• Fixed: on a Mac, the window moves when you drag anywhere on its top bar, the title included, instead of selecting the title's text.
• Fixed: a dictated message could include words you never said, made up during pauses and quiet moments. Pauses and silence are now
  cut out before your words are written down.
• Fixed: a long spoken answer stopped halfway on Linux and Windows. The voice now says it to the end.
• Fixed: the voice card of a chat you had left could stay in the sidebar for good, named "New session", with buttons that did
  nothing. It now goes when the voice ends, and the chat that is listening is never closed to make room for others.

Jauvex Personal 1.8.1

Jauvex Personal 1.8.1 for macOS (Apple silicon): open jauvex-personal-1.8.1-mac-arm64.dmg, drag Jauvex to Applications and open it. The app is signed and notarized by Apple.

On macOS or Linux you can also install it with one command: curl -fsSL https://jauvex.ai/install | sh. On Windows, in PowerShell: irm https://jauvex.ai/install/windows | iex.

SHA-256 of the disk image: a99748b8d564b88452020070f0aecaae943968b15cb06ad4c2220600497b88b8

What's new

• Up to date and safer: the app now runs on Electron 44.9.3, with the latest security fixes from Chromium and V8 and safer secure
  connections.
• Outside pages kept apart: pages opened in the right pane always run in their own sandbox, cut off from the app.
• Claude agents on the latest Claude Code (2.1.310), with its security fixes.
• Fresher building blocks: the app's other packages updated to their patched versions.

Jauvex Personal 1.8.0

Jauvex Personal 1.8.0 for macOS (Apple silicon): open jauvex-personal-1.8.0-mac-arm64.dmg, drag Jauvex to Applications and open it. The app is signed and notarized by Apple.

On macOS or Linux you can also install it with one command: curl -fsSL https://jauvex.ai/install | sh. On Windows, in PowerShell: irm https://jauvex.ai/install/windows | iex.

SHA-256 of the disk image: 7467153a2ffb9b90c8d209d27035b3452e80c84581a24ec450918e0de66e4896

What's new

• Open models: run an agent on a model of your own: Ollama, LM Studio or vLLM on your computer, or OpenRouter. Set it up in
  Accounts, where List models shows what the server offers, then pick Open models when you start an agent.
• A tidier left panel: each folder shows its agents, workflows and boards as a small tree, each group with its count and a plus to add
  one; three rows show, then "N more…", and a group folds from its name.
• Tabs and a filter in the left panel: All, Agents, Workflows and Boards, and a filter that narrows every folder as you type (Escape
  clears it). The tabs, the filter and the Jauvex agent stay in place while the folders scroll.
• Drag to resize: the line above a workflow's or a board's chat drags to make it taller or shorter, and the height is kept.
• Answers as buttons: when an agent asks you to choose, its options show as buttons under the question, the recommended one marked. A
  tap sends your answer, the pen lets you write your own, and Alt with a number picks one from the keyboard.
• Quotes you can click: a quote in your message shows as a quote above it, and a click on it takes you to the text it came from.
• Fixes: a long quote no longer squeezes the message box; the Jauvex agent's row stays lit while its chat is open.
• What an agent needs from you, in yellow: a question or a request for you stands out in its own box; a click on it starts your reply.
• Questions for you stay in sight: when an agent asks you something while answering another agent, the yellow box stays in your chat
  instead of folding away with their messages, and it is not sent to that agent.
• Reply to part of the chat: select any text in the chat to quote it in your message, or right-click for Copy and Quote.
• Copy code in one click: every code block has a copy button.
• Older agents learn the new rules: Codex and Grok agents made before an update get what changed in the app's rules with their next
  message, and Claude agents always run with the current ones.
• An agent's words are never shown as yours: a first message or a message an agent sends from its terminal shows as that agent's, so
  the agent receiving it never takes it for your words.
• Spoken orders that name a folder act at once: "make a Codex agent under webapp" makes it, without asking again.
• Pictures stay in the Jauvex agent's chat: a picture you sent it is still there when the chat opens again.
• Voice never mutes you mid-sentence: auto-mute waits while you are talking.
• Answered questions calm down: a yellow box you clicked, or that you replied after, is no longer highlighted.
• Clearer labels: "Full access (YOLO)", and Codex credits read as a number ("Credits: 62,500"). Settings says why an idle chat compacts a
  little earlier.
• Fixes: no "Send now" on a message waiting for a compaction; the help button always reaches the Jauvex agent; an update no longer counts
  the Jauvex agent itself as busy; a Codex sign-in refreshed elsewhere is picked up again; a YOLO Claude agent is never stuck on a permission
  card nobody sees.
• A lead is woken only when it asked to wait: the app no longer guesses a wait from an agent's words, and the wait order takes an
  agent's exact name or id, so it never binds a different agent with a similar name.
• Agents can't raise what agents may do: an agent's order can lower a provider's permission level, never raise it; only you can, in
  Settings, Safety.
• Safer downloads: the app no longer accepts Brotli-compressed answers from the web, which a flaw in Node.js made unsafe; link previews,
  open models and updates use gzip or nothing.
• Agents keep everything organized: every agent is told, first after security, to keep everything in order: files and folders, temp
  work and the processes it started, code, branches and versions, tasks, decisions, instructions, agents, workflows, messages, secrets and
  memory. No stray copies or leftovers in your folders.
• Arch Linux and Omarchy: sh scripts/install-arch.sh installs and builds the app on Arch and Omarchy (Hyprland on Wayland), with
  whisper-server, a jauvex command and an entry in the app launcher.
• A message with no answer says so: when an agent's reply is empty, or a turn is cut off or lost, the line under your message says
  " didn't reply", with Send again, instead of leaving it looking swallowed.
• The voice keeps talking: Kokoro, the spoken voice on Linux and Windows, no longer crashes after several minutes of speech.
• "Hmm" sounds like "hmm": the voice on Linux and Windows says "Mmm", "Hmm" and "Mm-hmm" as sounds instead of spelling them letter by letter.
• The voice comes back by itself: if Kokoro ever stops, it starts again on its own, so your chats are not left silent until you switch
  voice off and on.
• No voice servers left behind: restarting, updating or resetting the app now stops its speech servers too, so the next start can
  always listen.
• Messages between agents arrive whole: a message to another agent with a code block in it is no longer cut at that code block.
• Replies reach Codex agents: an agent addressed by a longer id with a hyphen in it, as Codex agents often are, is found again.
• No more "OUTCOME: done" at the end of answers: the line a workflow reads is left out of the chat, of what the voice says and of
  answers sent back to other agents.
• Updates on a copy in the making: a copy built from a version in the making (1.8.0-dev) is offered the release and greets you after
  an update, as any other copy does.
• Long dictations heard better: a long spoken message is split only at a pause between words, never in the middle of one, so names
  and phrases no longer come back garbled.
• Pictures with spaces in their path show: an agent's picture whose path has spaces, %20 or a Windows drive shows in the chat
  instead of raw text.
• Voice turns off when you say so: a chat never listens twice, so your words no longer arrive twice, and turning voice off while it
  starts really stops it.
• Never a blank window: if the window's page crashes, fails to load, hangs or comes up empty, it loads again by itself (your running
  agents carry on), and the debug panel says what happened.
• No more "JavaScript error" dialog when a program hangs up: when Codex or a sign-in closes while the app is still writing to it, the
  app carries on instead of showing a fatal error box.
• No more blinking text under the orb: with voice on, the newest line of an agent's answer no longer flickers half drawn on a Mac.
• The message box always fits: on a short window, with a picture attached or a long draft, the message box and its buttons stay whole
  at the bottom; the dashboard and the messages above it make room instead.
• Nothing spills out of the chat: long paths, links, words, pictures and tables wrap or shrink to fit the chat, and code blocks scroll
  inside themselves.
• The welcome screen fits a small window: at 900 by 600 everything on it shows without scrolling.
• Signing in to Grok works right away: after grok login, new Grok agents start without restarting the app.