Prosights.ai
AI Products

Release notes

Every release of Kyberna, Kyberna Write and the Kyberna Agent SDK, newest first.

1.0.1

November 13, 2026
Download Kyberna Kyberna-1.0.1.dmg · 238 MB SHA-256 8fb1309683e618f68accf06c6398e2660ec03365b725bb1f5c5109e3407efda2

The first official release: Kyberna, Kyberna Pro and Kyberna Ultra as editions of their own, each opened with the license key from your purchase email, updates that install themselves, folders and files you protect from agents, and search in Settings. First launch: open the disk image, drag Kyberna into Applications and open it, confirming once when macOS asks; the setup sheet installs the gateway itself.

New

  • The license key, asked for once: on a copy of Kyberna that asks for a license key, setup's first page takes the key from your purchase email and names your edition and the day you bought it; on a Mac where setup has run, a sheet of its own asks instead. Settings › General shows the key after your Mac's login, with Copy for your other Mac, and an Upgrade row for a higher edition's key. No copy asks for a key yet.
  • Meeting transcripts reach a cloud model with placeholders: with Protect personal information on, the names and numbers Kyberna finds in a transcript, the meeting's title and the attendees' names in the forms people say them are replaced before a model in the cloud reads them, also after Change… and when a model runs out. The notes show the real values, underlined, and a new Meetings row in Settings › Privacy chooses this per place. If Kyberna can't check the transcript, it sends nothing.
  • Kyberna and Kyberna Write update themselves: each app looks for a new version once a day, downloads it in the background and installs it the next time you quit and open the app. Settings has Updates with 2 switches and Check Now, and each app's menu has Check for Updates…. This version is the last you install by hand.
  • Find a setting by its name: typing in the Search field at the top of the Settings sidebar now lists each matching setting with its pane's name beneath it, as System Settings does, and choosing one opens that pane scrolled to the setting. Your search stays in the field, and clearing it lists every pane again. Before, the search narrowed the sidebar to the panes that matched and opened a pane at its top.
  • Protect your own folders and files from agents: in Settings › Folder Access, under Protected Folders and Files, click Add Folder… or Add File… and choose what agents may do with each: No Delete (they can change it but can’t delete or move it) or Read Only for a folder, and No Delete, No Change or Read Only for a file. Every agent Kyberna runs is held to it, and so is a command you type after "!"; the conversation says which folder or file Kyberna stopped the agent from changing. Documents, Desktop, Downloads, iCloud Drive, the Applications folders and your home folder are listed at the top at No Delete. In Kyberna Pro and Ultra.
  • Kyberna comes in editions: Kyberna and Kyberna Pro: Kyberna Pro has every feature; Kyberna leaves out Channels with Slack and Signal and their guides, Rogue Agent Protection, Permission Boundary, Protect personal information, Folder Snapshots, Meeting Notes with its action items, and serving Kyberna's local models to other Macs, and its Settings, About window and guide show only what it has. Kyberna Ultra is coming later.

Changed

  • Fewer questions for steps you already asked for: a site you write in your request as a plain domain (linkedin.com) now counts as the site you named, as a full address already did, so opening it asks nothing; a file name such as main.py or notes.md is not read as a site. After you answer Allow to a step, the identical step and the same kind of command no longer ask again in that conversation, and another page on a site you allowed opens without a question; a command that reaches the network, installs a package or deletes, a send, and a write outside the conversation's folders ask every time. When you add a folder, in Add Files and Folders… or with Add the Folder on a question, Access: lets you choose Read & Write or Read only; an engine reads a Read only folder without asking and asks before each change there, at every approval level. Background agents follow the same answers. On the 159 questions measured on 2026-11-10 this removes 37. Edition: Kyberna. (guard 50, v0.3.3-PM.100017)
  • Kyberna never lets an agent delete your main folders: a delete, move or rename of Documents, Desktop, Downloads, iCloud Drive, the Applications folders or your home folder is refused at once, with no question, whether or not the folder is in the conversation and whether Dangerous Delete Guard is on or off, also when the folder is named only as the command runs or inside a script it runs; the conversation says which folder Kyberna stopped the agent from changing. A command you type after "!" that would delete or move one of them does not run. Running a command yourself with "!" is now in Kyberna Pro and Ultra, and in Kyberna the line says so.
  • Folder Access names macOS's permission and lists iCloud Drive only when you use it: Settings › Folder Access heads its list "macOS’s permission for Kyberna Engines", and the iCloud Drive row shows only when iCloud Drive is turned on on this Mac.
  • The Settings window is titled with the pane you are in: the title reads General, Models, Tool Policy and so on, as System Settings does, in place of "Kyberna Settings" on every pane.
  • Tool Policy's column headings stay in view: in the Kyberna tools table, the rows scroll under the headings Tool, Console, Channels, Reminders, Tasks and Agents, and searching Settings for a tool's name, such as Shell commands, opens Tool Policy at that row.
  • Reminders run on any model you have: Settings › Reminders › Model lists the same models as Meeting Notes' Write notes with: every model of your accounts and your local models. With nothing chosen, reminders run on the Console's default model; reminders set up before this update keep running where they ran until you choose another model.
  • Settings › Models lists each engine under its vendor: after Anthropic and Google come OpenAI, xAI and GitHub Copilot, each with its sign-in and switch, and the rows read Subscription (Antigravity), Subscription (Codex), Subscription (Grok Build) and Subscription (GitHub Copilot). The engine picker groups its rows under the same vendor names.
  • The xAI switch says what it protects: it reads Stop xAI from keeping my coding data, is on while xAI keeps nothing, and lines up with the other switches. Your xAI account's setting is unchanged.
  • Choose what Antigravity CLI may do in Slack and Signal: Settings › Tool Policy has an Antigravity CLI section with Run command, Write file and Edit file under Channels. Each starts at Deny, so Antigravity CLI runs no command and changes no file in a channel's folder until you choose Allow.
  • The delete question no longer offers a Delete that cannot delete: a delete on a script's line, through a shell variable, or joined to another command in the same line stays even after you answer, so the sheet now offers Run Script or Run Command and says the delete will fail. Asking the agent for that delete on its own gets a real Delete.
  • Restore covers every folder of a conversation: a turn's snapshot used to hold only the working folder, so what an agent changed in a conversation's other folders could not be put back. Now each folder has its snapshot before the turn's first change, and the turn keeps 1 row and 1 Restore…, whose sheet names each file's folder. Undo Restore puts every folder back. A folder Kyberna cannot take shows its own line, and for a large folder you are asked once. See Conversations.
  • Dates of birth and ID numbers are hidden without a label: with Protect personal information on, a date of birth, a Social Security number written without dashes, a passport, driver's license, National Insurance, employer ID, bank account or routing number is now replaced before a cloud model reads it when nothing beside it says what it is, as long as its column, its key or another personal detail in the same record does. A 10-digit account number is hidden as a bank account, not a phone number. See Privacy.
  • Names are found by Kyberna Redaction, which comes with Kyberna: with Protect personal information on, Kyberna's own name model, which runs on this Mac, finds the names it replaces; in Kyberna's test it found 213 of 214 names. Nothing is downloaded for it, so Settings › Privacy no longer shows a Name detection model row. See Privacy.
  • Your conversations' key moves into the Keychain: on this Mac the key that encrypts your conversations moves to the Keychain the first time the new version starts, with nothing for you to do. A program running under your login can no longer read the key's files and open the record file with them. It can still get conversation text by running kyb while Kyberna is running.
  • Antigravity CLI keeps its own permission checks on: Kyberna no longer turns off agy's checks. agy reads files, runs commands, writes files and reads web pages only through Kyberna's tools, its own sub-agents are off, and when its checks do not load in a conversation Kyberna stops it and the conversation says so in 1 line. Its other 7 tools are checked only by Kyberna's tool policy. See Use Antigravity CLI.
  • A failed update check says what to do next: when a check you started cannot finish, Kyberna and Kyberna Write say what happened and what to do next, with OK, in place of "Update Error!" and Cancel Update.
  • Other people's words stay out of your lines: in Meeting Notes, the short things others say through your speakers, such as "Thank you.", show under Others and not under You, also after a pause. Your own short lines stay.
  • Kyberna Guard lists Decide for me: Settings › Kyberna Guard now shows Decide for me with its switch beside the other protections. The details under an error in Decide for me and under a failed Name detection download open with VoiceOver too, and stopping a turn while Kyberna reads a folder's access no longer crashes.
  • Kyberna reads a shell script before it runs: under Decide for me, a script line that deletes outside the conversation's folders, or deletes a path known only when it runs, now comes to you instead of running. The approval sheet and the delete question name the script and its line: "Line 3 of “cleanup.sh” deletes outside this conversation’s folders." See Tool policy.
  • The delete question says what a path holds: "Qwen3.6 35B wants to delete the folder “old-site” in Documents › Sites and everything in it" comes with where it sits against your project and how many items it holds, 214 for example, of which 180 were there before the conversation started. Show Command lists each full path. See Tool policy.
  • Every engine menu lists engines the same way: the composer, the start page, New Task, Settings › Models' Default model, a channel's Engine, Reminders' Model, Meeting Notes' Write notes with and the blocked turn's Another engine list the same engines under the same vendor headings as Settings › Models.
  • Settings no longer wait on a slow read: the record menu shows Follow the default at once and adds its period, such as (week), once it is read; the gateway waits at most 2 s for the retention setting and keeps 30 days if it cannot read it.
  • Each feature names the research behind it: the guide's Research page and the website's features page list the sources each feature rests on, with each paper's verdict and its reason; Gemini CLI is listed there under Retired features.

Fixed

  • Grok Build no longer runs your Claude Code or Cursor settings: a Grok Build conversation starts with Grok Build's Claude Code and Cursor compatibility off, so your Claude Code skills, rules, instructions, MCP servers and hooks no longer run inside Grok Build without Kyberna asking you first.
  • A file an agent reads names itself: when an agent reads 2 files at once, each result opens with the name of its own file, so neither can be taken for the other, on every engine Kyberna reads files for.
  • A reply that stops arriving ends as stalled: when a model reached through an API stops sending its reply, Kyberna ends the turn as stalled, also while your Mac is busy, in place of showing it as finished.
  • A shell command sent cut short is refused before it runs: when an agent sends an incomplete shell command, Kyberna refuses it before it runs or asks you, with 1 sentence asking the agent for the whole command.
  • Protect personal information keeps the real values where they belong: with Protect personal information on, GitHub Copilot CLI on a channel writes a file with the real values instead of placeholders, a write Kyberna refuses is shown without placeholder markers, the same name written in different case or spacing ("Ana Ruiz", "ana ruiz") gets 1 placeholder, Copilot CLI's "Disabled tools" notice no longer shows in the conversation, and hiding personal information in a long result no longer slows as the conversation collects more values.
  • Kyberna Guard lists Protected Folders and Files: Settings › Kyberna Guard now shows Protected Folders and Files beside the other protections.
  • Leave Out of Context shows only in Kyberna Ultra: in Kyberna and Kyberna Pro, the transcript's Leave Out of Context command and the Tasks pane's line about what a run reads no longer appear, since those editions have no Token Optimization. Edition: Kyberna Ultra.
  • Tool Policy's table keeps its columns side by side in a narrower window: in Kyberna, which has no Channels column, the table now stays in its wide form down to the width its 4 columns need, in place of the width 5 columns need. Edition: Kyberna.
  • Kyberna's guide links to no page it leaves out: in Kyberna, the guide's lines about channels, Slack and meeting notes now name Kyberna Pro and Kyberna Ultra and link to no page this edition does not have. Edition: Kyberna.
  • Kyberna Guard and Help describe Dangerous Delete Guard alike: both now say it refuses a delete of the top of the disk, a folder at its top, your home folder, Documents, Desktop, Downloads, the Applications folder in your home folder, a conversation's folder, this Mac's temporary folder, a folder that holds any of them, or iCloud Drive itself. Edition: Kyberna Pro.

0.3.3

November 9, 2026
Download for Mac Kyberna-0.3.3.dmg · 121 MB SHA-256 697867d99aeeebe835cdcbb9615e616865cf1f864eb62d0588287a9d0ab658e8

Agents you can start and come back to, a command line in the composer, attachments, 30 days in Recently Deleted, and models on this Mac that read your images and PDFs. First launch: open the disk image, drag Kyberna into Applications and open it, confirming once when macOS asks; the setup sheet installs the gateway itself.

New

  • Run a command yourself from the composer: start a line with ! and press Return, as in !git status. It runs on your Mac as you, in the conversation’s folder, with no model turn, and a row shows the output. The output goes with your next message unless you click Leave Out, and a command that names a protected place keeps its output from the agent. To send the words instead, click Send as Message.
  • Local models, model servers and the Anthropic API read images and PDFs in your folders: ask for "report.pdf" or a screenshot and the model reads it, under the same folder rule as a text file. A PDF comes as text, up to 20 pages at a time, with scanned pages read by text recognition on your Mac; a model that does not see pictures, or a server or API that refuses one, gets the words in an image and says the picture was not shown, and the conversation goes on.
  • Reach the top and the bottom of a long conversation in 1 action: when you scroll up, Jump to Latest appears above the composer and returns you to the newest message. View › Scroll to Top (⌘Home) and View › Scroll to Bottom (⌘End) do the same from anywhere in the conversation, the composer included.
  • A message that starts with @ reaches the other conversation with no model turn: "@Release notes what is the time" goes straight to that conversation, and "ask @Alpha for the status and @Bravo for the plan" gives each conversation its own part. A question about a conversation, such as "what did @Alpha decide", still goes to the model. Twins with the same title keep the one you picked, and "project/title" typed by hand works when it can only mean 1 conversation. A message to a conversation whose turn fails now shows Not answered with the reason, and a local model waiting its turn no longer ends as "sent nothing for 300 seconds".
  • Ask the agent about a request before you answer it: an approval request and a step Rogue Agent Protection holds now have Ask About This Request…. The sheet closes without answering, the request waits at the foot of the conversation with the same buttons, and a panel at the right of the window lets you ask the agent questions about it. The agent answers from the conversation without running anything. While you have typed something in the panel, the decline reads Don’t Allow and Send: it refuses the request and sends your words to the agent in place of the usual "stop and ask" answer, and the transcript shows what you wrote.
  • Decide for me answers your agents’ questions for you: in Settings › Tool Policy › Approvals, the new level lets Kyberna’s checks answer the questions an agent’s calls would ask, as Claude Code’s auto mode, Codex’s auto-review and Antigravity CLI’s auto mode do. Reads and file edits in the conversation’s folders run, a call the checks find safe runs, and a call they find unsafe is refused and the agent is told why at once. Kyberna asks you about a call the checks find risky or cannot judge, the first read outside the conversation’s folders and any call your own rules cover; after 3 refusals in a row or 20 in a conversation it asks you again until you allow a call. Under each reply, a line lists what Kyberna allowed and refused for you.
  • The guided browser’s side panel says when Kyberna hid personal information from the model: under "What Kyberna did", a row after the page read says what Kyberna hid, for example "Kyberna hid 1 email address in this result from the model."
  • GitHub Copilot CLI and Grok Build read your files through Kyberna’s tools while Protect personal information is on: like Codex and Antigravity CLI, they read, search and run commands through Kyberna’s tools, so their models see placeholders in place of the personal information in your files, and Settings › Privacy no longer says these agents read files with tools Kyberna cannot change. GitHub Copilot CLI’s note listing the tools it was started without no longer appears ahead of its answer in a conversation.
  • The model picker lists Antigravity CLI’s models: under Antigravity CLI’s row, each model your Antigravity CLI account offers, with its name and its id, the newest of each family at the top and the rest under More Models; the one you choose runs the conversation, and a change takes effect on the next turn of the same conversation. Before, only Antigravity CLI’s own row showed, so every conversation ran its default model.
  • Rogue Agent Protection holds a step that would send your details to a website: when an agent tries to open a web address that carries an email address, phone number or street address it read in the same reply (from Contacts, Mail, Notes, a page or a file), or to type or send one to a website you did not name, Kyberna holds that step before it runs; the card names the kind of detail, where Kyberna read it and the website, never the detail itself, and you choose Stop or Continue. This also covers the web fetches of Codex, GitHub Copilot CLI, Grok Build and Antigravity CLI.
  • Your MCP servers reach the engines you use, and every call to them goes through Kyberna’s approval: Claude Code, Copilot, Grok Build, Codex and Antigravity CLI can use the MCP servers you added in Settings › Tool Policy in conversations, and so can conversations on local models, model servers and the Anthropic API while Kyberna’s tools are on for them; reminder and task runs get the servers with every call deferred, and channels get none. Kyberna starts the server itself, offers the engine only the tools you accepted or that were recorded clean, checks a tool’s text again at its first call, and forwards the call only after your policy and you allowed it; the Settings footer and the guide’s Tool Policy page say who gets them.
  • A server’s tools are listed again at the start of a conversation when the last listing failed: a server whose tools Kyberna could not list, or whose flagged text you have not accepted, stays out of that conversation and the conversation notes it, and Kyberna lists it once in the conversation’s folder, for up to 30 seconds, and records the result in the audit. While you turn a server on, its row reads "Starting it once to read what its tools say about themselves. This can take up to 30 seconds."
  • The guide has a page, The research behind Kyberna: it lists the psychology and behavioral science papers Kyberna’s features are designed on, 2 so far, each entry giving the paper’s authors, venue, design, sample and finding and the feature it shaped. A paper is listed only after its method and results were read in the paper itself; 188 further citations are waiting or were left out, each with its reason in the repository.
  • Rogue Agent Protection asks the agent to rewrite a delete whose folder is known only when it runs: a command such as rm -rf $(dirname $d) or rm -rf "$L/" with L not set in the command is refused before it runs, and the agent is asked to name the folder’s full path and run the delete again; the transcript shows 1 line for each, and nothing waits for you. A 3rd in the same reply, and any delete of the top of the disk, your home folder or a conversation’s folder, waits for your answer on Rogue Agent Protection’s message. kyb gateway policy revise off makes every such delete wait for your answer instead. With GitHub Copilot CLI, Codex or Grok Build the agent is asked as its next message, once its reply ends, and the 2 rewrites count across both replies.
  • Rogue Agent Protection’s message explains the step it holds: under what the agent said, Explain this request opens the same plain sentences as the approval sheet: what the step does and what it reaches, why Kyberna rates it as it does, and Show Command with the exact command. For a delete whose folder is known only when the command runs, such as rm -rf $(dirname $d), it says the folder deleted is the one that holds $d, and that which folder that is cannot be known before the command runs; the approval sheet says the same for that command. A command such as $(dirname $d) stays on 1 line. The explanation opens by itself when Folder Boundary marks the step, scrolls inside the message in a small window, and opening it answers nothing.
  • On a release, your sync passphrase and channel secrets stay in this Mac’s Keychain: the gateway moves them there once, reads each one back, and only then deletes the earlier copy; Kyberna reads them without asking you. If Kyberna cannot read one, it says which to enter again and makes nothing in its place.
  • Start agents, leave, and come back to their results: on Claude Code, the external agents and Antigravity, as on a model on this Mac, a model server and the Anthropic API, a conversation replies when its agents finish or fail, with nothing typed, and a conversation already working takes the result in that work or right after it. Agents that end together get 1 reply, and a paused conversation replies once you resume it. See Agents a conversation starts
  • A stuck agent is restarted: an agent that writes nothing for 10 minutes, or 16 on Claude Code, the external agents and Antigravity, is sent its task again, up to 5 attempts, and the row says so: "It was stuck for 10 minutes, so Kyberna restarted it (attempt 2 of 5)." After the last attempt the conversation is told what the agent did and replies. See Agents a conversation starts
  • The conversation’s model knows what you did to an agent: it is told when you pause, resume or stop an agent, and a stop passes on what the agent had written. A kill in the agent’s own window now reads "You stopped it." See Agents a conversation starts
  • A model no longer reads a line from Kyberna as your message: an agent’s end, or a message from another conversation, that reaches a running turn says who sent it and that you did not write it, on every engine.
  • Attach files from the composer: click the paperclip (Attach files) beside the send arrow, paste an image or drop files on the composer to add up to 20 images, PDFs and text or code files to a message. Each file shows above the prompt field with its name and size. The engine receives what Kyberna read from the file when you chose it, never the folder it is in, so the files beside it stay out of the conversation. An engine that does not read pictures gets your message, and the conversation says the picture was not sent.
  • Deleted conversations, tasks and meetings wait 30 days in Recently Deleted: Delete in the sidebar no longer asks. The item leaves the sidebar, and View › Recently Deleted lists it, newest first, with the days left. Select items and click Recover to put them back at once, or click Delete Immediately… to delete them for good. A Private conversation keeps its question and is deleted for good.
  • A turn’s tool calls fold under 1 heading: the calls a turn makes, such as reading files, searching or sending mail, appear under 1 heading in plain words with counts, and open to 1 row per call, each titled in plain words, in place of the tool’s name and arguments. A turn with 1 call keeps its row.
  • Kyberna refuses a delete of a place that holds your work, without asking: a shell delete of the top of the disk, your home folder, a conversation’s folder, a folder that holds one or this Mac’s temporary folder is refused before it runs, with no question first. The line in the conversation names the folder, for example “Kyberna stopped Grok Build from deleting “~/Projects”, which holds this conversation’s folder. Nothing was deleted.”
  • Kyberna’s own engines never read a page from your own network: a web address that is, or resolves to, an address on this Mac or on a private network is refused before any request, and so is a redirect into one. The model reads 1 sentence saying why, for example “192.168.1.1 is an address on a private network, so web_fetch did not read it; web_fetch reads public web pages only.” A name that answers with a public address first and a private one second reads nothing from the private one.
  • A page longer than about 19,000 characters comes in parts: on a local model, a model server and the Anthropic API, the model reads a long page one part at a time and asks for the next when it needs it. Each part ends with a line saying how much is unread. Before, the model got the first part and no way to read on.
  • Undo a turn’s changes in Documents, Desktop, Downloads, iCloud Drive and cloud storage folders: a conversation in one of these places gets the snapshot and the Restore… row once macOS has allowed Kyberna Engines into it. Until then the turn says in 1 line that it ran without a snapshot, and deletes in the folder ask first. Click Allow… next to the place in Kyberna › Settings › Folder Access.
  • A command’s connection to a website follows the Network level on every engine: in Settings › Tool Policy › Approvals, Ask every time asks about every website, Ask only for risky calls reaches a website you wrote or a web search returned without a question and asks about any other, and Decide for me judges the website as it judges a page an agent fetches. Each connection made without a question is recorded in the audit log. A reminder or a task follows the same levels, and Approvals shows a Tasks row.
  • An agent can look up your own name, email address and phone number: asked for “me”, a lookup in Contacts answers with your My Card, under the same Contacts permission as any other lookup. With no My Card it says so, and the model is told not to invent them.
  • Add single files to a project, not only folders: right-click the project’s name in the sidebar and choose Add Files and Folders…, or use the button under the list in the “Folders and Files in” window. Each file you choose becomes a row of its own, an engine reads that file and not the files beside it, and no conversation moves.

Changed

  • A model’s reasoning is folded under its heading until you open it: a reasoning row shows only "Thought for N s" ("Thinking…" while the model is still reasoning) with a chevron at its end; a click on the heading opens or folds the reasoning, also while it streams, and what you chose holds after it ends. (console 547)
  • More of what someone else wrote reaches the model marked as data, on every engine: the lists a mail or notes search finds, a web page’s text in Kyberna’s browser, text of 400 characters or more you paste, the text you select in Kyberna Write and ask about, a reminder on a list other people share with you, and on Claude Code what your MCP servers and the browser return now reach the model between the markers that say it is data, not instructions, with hidden characters removed. A reminder on a shared list is no longer introduced to the model as your own. Your conversation shows what you pasted or selected as you sent it.
  • A file you ask for inside the conversation’s folders is written without a question, on every engine: in a Console conversation, when the engine writes or changes a file inside the folders you gave the conversation, Kyberna no longer asks first; Kyberna’s own engines (the local models, a model server and the Anthropic API) now write such a file with a file tool of their own instead of a command that asked. Kyberna still asks before a write outside those folders, a write to a file holding keys or startup or agent settings, a write to a file whose name came from something the engine read rather than from you, and any delete.
  • Add the Folder on the question asked with Folder Boundary off: when Folder Boundary is off for an engine and a request reads, changes or runs a command on a file in a folder outside the conversation’s folders, the question offers Add the Folder in place of Allow for This Conversation. It adds the folder to the conversation (to the project, for a conversation in a project), keeps what Allow for This Conversation would have kept for that request, and runs it; its line says both. In a place macOS protects, such as Documents, it first shows what macOS asks next.
  • Meetings move to your new Mac with your conversations: a move file made with Make Move File… now carries your Kept meetings, their transcripts and notes, encrypted, and Choose File… adds them on the new Mac. A meeting still being recorded moves as it stands, the file’s line names it, and the new Mac shows it as ended. Private meetings are not moved.
  • Antigravity CLI’s usage logger is refused: Kyberna’s network gate refuses Antigravity CLI’s connections to play.googleapis.com, Google’s usage logger, without asking you, and lets every other connection through; Antigravity CLI therefore runs only where the Kyberna Gateway runs, and the Console’s direct mode and kyb refuse it as they refuse every agent behind the gate.
  • Antigravity CLI can run on your Gemini API key: in Settings › Models › Google and on the setup sheet, Antigravity CLI’s row lets you choose "Gemini API key" or "Google sign-in". On the key, Kyberna gives Antigravity CLI the same key the Gemini API uses, writes it into no file of Antigravity CLI’s, and never reads your Google sign-in; Kyberna asks before you switch from the key to the sign-in, and never switches ways on its own.
  • Kyberna’s requests leave LM Studio’s log after each turn: LM Studio writes what its models read into its own log on your Mac; after each turn on an LM Studio model, Kyberna removes its own requests from that day’s log and leaves everything else in it, and Settings › Models says so under LM Studio’s row, with what LM Studio keeps while a model stays loaded.
  • The conversation says when an external agent did not load Kyberna’s tools: when Kyberna’s tools are on and the agent never loads them in its first turn, 1 note at the end of that turn says the agent cannot use them and that a request needing one is not done, so a missing email or note is not taken for done.
  • An agent stopped at its limit says so in the conversation: when an agent you started runs out of model calls, its row reads "It couldn’t finish: it reached its limit of 40 model calls before it finished." instead of the last thing the agent wrote.
  • An agent’s own words in its failure row read as it wrote them: when an agent you started fails with its own words, its row shows them after "It couldn’t finish: " with their first letter as written, so a code or a name such as "FOUND-PART-1" no longer reads "fOUND-PART-1".
  • An agent has the same Kyberna tools as the conversation that started it: when you turn Kyberna’s tools on for a conversation, every agent it starts afterward gets Mail, Notes and the rest too, on every engine, and when you turn them off its agents get none, even on Claude Code and the Anthropic API; an agent’s agents follow the same switch. An agent already running keeps what it started with. Every email an agent sends still asks you first, naming the agent. Before, an agent took its engine’s default whatever you had set for the conversation.
  • The model picker shows the newest model of each family: for the Anthropic API, the Gemini API and GitHub Copilot CLI, every other model is under More Models at the end of that list, open at once when your conversation is on one of them. The Anthropic API rows are the models the API lists for your key, for example Claude Opus 5.5; before, the picker offered 3 models written into Kyberna.
  • Kyberna reads the text inside an MCP tool’s input fields too: when a tool’s description or title in any input field tells the model what to do, the turn-on refusal, the Settings row, kyb mcp list and the conversation-start note quote that text and name the input it was found in, for example "the description of the “query” input of find tells the model what to do …"; a tool whose input text waits for your review is left out with the line "find (the text of its input waits for your review)", and a tool you accepted before inputs were recorded is shown once more.
  • Settings › Kyberna Guard lists Telemetry Off and Kill All Activity as rows of their own: Pause All Activity says pause and resume, the pause and kill of one conversation is named Pause and Kill a Session, Confirm it is you is 1 sentence, and the Kyberna Guard guide page lists the same rows and sentences.
  • Confirm it is you says what holds: a confirmation covers 5 minutes of the actions taken in the same place, so a person who hands a conversation to a terminal and then copies it is asked twice; Kyberna Guard’s row and the guide say so.
  • A held run’s row in the Gateway’s menu is its name over the sheet’s sentence: for example Task “Nightly dependency check” over "Kyberna is 92% sure Claude Code went off track", with no “ · 92%” after the name, and "Claude Code is asking for approval number 6 in this run" for a run held at its approval limit; a name longer than 40 characters ends on a whole word with "…".
  • A loaded model says whether it is in use: its line reads "in use" only while a conversation holding it has a turn running, "not in use" while conversations hold it but none is answering (Unload frees it), and "idle" with the minutes once none holds it; conversations show the names the sidebar shows, and the others are counted. Before, a model read "Peru, agent:b, 1 unnamed lease · loaded 1 min ago · in use" whenever anything held it; now it reads "Peru, 2 other conversations · loaded 1 min ago · not in use". kyb models loaded prints the same words.
  • The memory check counts the memory LM Studio and Ollama hold once: quitting a runner adds its footprint to the memory available, and the model’s need no longer also counts memory the runner holds. Before, a model could read "does not fit" while it fit; on the figures of 2026-10-11, 7 such answers become "tight" or "fits", and kyb models check prints, for example, "verdict: fits (need 2.4 GB against 35.4 GB available)".
  • A conversation waiting on Kyberna’s model server says so: on the server engine, while the model answers another request, the status line reads "Waiting for the model: it is answering another request." until the model’s first output replaces it.
  • The record’s facts row for the Gemini API engine names the free tier: it says "On Google’s free tier, people at Google may read what you send, and Google uses it to improve its products." before the paid key’s sentence, and the Privacy guide page says the same and adds Google’s own request, "Do not submit sensitive, confidential, or personal information."
  • Antigravity CLI’s terms line states what Google’s FAQ says: Settings, the setup sheet and its info button, and the refusal when Antigravity is off read that it is not supported on consumer plans inside other applications, that other software accessing it breaks the Terms of Service and can get an account suspended or closed, and Google’s key route; the Engines guide page matches.
  • A secrets file in your folders is refused on more agents: with an agent’s switch on, a .env file (and .env.* except .env.example) inside the conversation’s folders is refused for reading and writing on Copilot, Grok Build and Antigravity CLI, and the conversation shows the protected-file line; with Antigravity CLI’s switch off it runs inside the protected places’ rules as the others do. Kyberna’s own file tools answer "Refused: <path> is a protected file (keys, passwords or another app’s private data); Kyberna never gives an agent access to it." and skip the file in searches. Codex with its switch on still reads a .env inside its folders.
  • Your cloud accounts’ keys are protected as ~/.ssh is: ~/.aws, ~/.config/gcloud and ~/.azure are refused to Claude Code, Codex, Copilot, Grok Build and Kyberna’s own engines, with an agent’s switch on or off, and the conversation shows the protected-file line wherever it shows one for ~/.ssh; the sheet reads "This folder holds your keys." Codex refuses them without telling Kyberna, as it does ~/.ssh.
  • A command Codex runs can no longer rewrite Codex’s own settings: its config.toml and AGENTS.md in the folder Codex runs from are refused to its commands, with Codex’s switch on or off, as Codex refuses them when it runs alone.
  • With Claude Code’s Permission Boundary off, a delete you allow outside the conversation’s folders now completes: when Claude Code’s sandbox refuses the command, Claude Code asks again to run it outside the sandbox, and Allow runs it, as Claude Code does on its own with its sandbox on. A folder you add to the conversation is writable from the next message. The network question and the protected places stay.
  • The folder pill’s last 2 items say what they do: "Add a Folder…" opens the panel that adds a folder, and "Show All, Remove, Add Folders" opens the window that lists, removes and adds folders and files; the project’s name no longer widens the item.
  • Gemini CLI is not in this version: the model picker, Settings › Models, the setup sheet, Accounts and keys, kyb agents and kyb engines do not offer it. A conversation you had on it stays in the sidebar and can be read, and sending in it says the agent is not available. Your Gemini API key stays with Google’s Gemini API, and kyb agents key gemini stores it there.
  • The New Task sheet picks the model as a conversation does: its Model is the composer’s engine pill, 1 control that shows the engine and the model together, with each engine’s models under its name, the newest of each family first and the older ones under More Models, in place of the sheet’s separate pop-up. A saved task keeps its engine and model.
  • Every place that picks a model lists the same engines: the start page’s engine pill, Settings › Models › Default model and a channel’s Engine list what the composer’s engine pill lists, each engine’s models under its name and the older ones under More Models. Default model also lists the model servers you added and those Kyberna found.
  • A meeting’s notes conversation shows the notes: Show Conversation on a meeting opens the conversation that wrote its notes with the notes drawn under the meeting window’s headings, Kyberna’s instruction as 1 line, and the instruction and the raw answer behind 1 disclosure.
  • The context ring shows in every conversation Kyberna Gateway holds: the ring marked ctx at the foot of a conversation shows how full the model’s context is in a conversation that continues when you close the window as well as in one this window holds. Antigravity CLI reports no use and shows no ring.
  • Code inside a sentence is blue under the Accent look: in the transcript, code and commands inside a sentence are drawn in the blue of Kyberna’s icon on a light blue shading, in Light and Dark Mode. The Classic look keeps the text’s color on a gray shading.
  • More destructive commands ask at Ask only for risky calls: a git push that overwrites, deletes or mirrors branches on the server, a git checkout or git restore that discards uncommitted changes, git gc --prune=now, swift -e and brew uninstall --zap now ask, and the sheet says why. brew uninstall without --zap keeps the app’s data and stays at medium risk.
  • One notice says why a conversation’s engine cannot answer, with the actions that fix it: when the engine cannot run the next turn because it is not available in this version, is turned off in Settings, is out of quota or its sign-in expired, your message waits and the transcript names the engine and the reason, with Sign in again for an expired sign-in and Use another engine for every cause. It replaces the separate sign-in and API-key banners.

Fixed

  • Slack’s sign-in opens in Kyberna’s browser: Slack no longer answers "your browser is not supported". The browser window, a private window and the Slack setup guide’s window now identify as the Safari on your Mac, so Slack shows its workspace page, and the setup guide’s first step opens it with 1 move in "What Kyberna did".
  • A mail with a recipient written in the wrong form is not sent: when a model writes a To, Cc or Bcc as text, as a list inside the list or with brackets or quotes around the address, Kyberna sends and saves nothing and tells the model which field to write again; a field holding 1 address written as text is sent to that address, and the send and draft sheets now show a Cc or Bcc written that way.
  • A local conversation at a small context size keeps answering after a restart: the conversation carried into the model after a restart now fits half the context size you chose; at 8K it was up to twice the size the model could read.
  • Kyberna’s summary of a long conversation is kept after a restart or an engine change: in a conversation on the API, a model server or a local model that Kyberna compacted, the model after a restart or a change of engine is given that summary, as it is given Claude Code’s, so it still knows the earliest part when the oldest messages no longer fit.
  • An answer cut off when Kyberna restarted is now said to be cut: when Kyberna stops while a model is answering, the conversation shows "The answer to your previous message stopped when Kyberna restarted and was not kept. To get an answer, send the message again.", and the model answering your next message is told that question was not answered rather than answered.
  • A conversation on a model server no longer says "Running on …" when it starts where you chose: after you choose a model on a server in the engine picker, the conversation shows that row only when it starts on another server or model than the one you chose.
  • A conversation on a model server you added names that server by the name you gave it: its notes in the transcript, such as the line when a turn could not be answered, say "Studio: glm-4.6" for a server you named Studio, instead of "custom server: glm-4.6". A server you added without a name keeps "custom server".
  • A notice’s explanation lines up under its title: in a conversation, the gray sentence under a notice’s title, such as the one after a high-risk command ran without asking, now starts every line at the same place, under the title, instead of starting its first line under the symbol and the lines after it further right.
  • A page that stops while it loads no longer leaves a browsing conversation waiting: when the browser’s page process ends during a load, the conversation now gets "the page stopped while it was loading; opening it again loads it afresh" within a second and can open the page again, instead of waiting with no answer.
  • A long command, change or list of items scrolls inside its sheet: when Rogue Agent Protection holds a step or Kyberna asks about a delete, a long command, a long change or a long list of items to delete now scrolls inside its own box, and the part under the title scrolls when the window is short, so the sheet stays inside the window and its buttons stay where they are, instead of the sheet growing past the screen.
  • A message too long for a local model no longer stops the conversation: on a Core AI model, a message larger than the model’s whole context window now ends its turn with 1 line, such as "This message is too long for the model: with it the request comes to 7,365 tokens and the model reads at most 4,096, so the model did not read it; send a shorter part of it, or choose a model with a larger context window.", and your next message is answered, instead of every later message failing with "invalid response: the prompt is 7365 tokens and the model’s window is 4096". The same holds on the other local runtimes, model servers and the API when a message cannot fit.
  • The row after Don’t Allow says what you declined: it now reads "You didn’t allow Claude Code to send an email to ann@example.com.", in the approval sheet’s own words, instead of a tool name and the instruction Kyberna gives the agent; a folder question you refused names the folder, and a refusal by Kyberna’s rules reads "Kyberna didn’t let Claude Code run a command: …" with its reason. Ask Kyberna shows the new row.
  • Explain this request shows a command’s substitution as the command wrote it: for a delete such as rm -rf $(dirname $d), the last sentence now reads "Kyberna cannot describe part of this request: the path $(dirname $d) is built when the command runs." instead of naming an internal placeholder, and 2 such paths in 1 delete are named in 1 sentence.
  • Explain this request says what a redirect with no command does: for > ~/.zshrc or : > notes.txt, it now reads "It empties ~/.zshrc, or creates it empty if there is no such file." instead of "Kyberna has no facts about this request."; >>, which adds to a file, is described as before.
  • The approval request names the whole file a command writes when its path is built when it runs: for echo hi > $(dirname $d)/x, the reason now reads "It writes the file $(dirname $d)/x." instead of "It writes the file $(dirname.", for >>, backticks and 2 substitutions in 1 path too.
  • Rogue Agent Protection no longer says a delete is outside the conversation’s folders when nobody can know its folder yet: when it holds a delete such as rm -rf $(dirname $d), the message now reads "Kyberna held this step because what Claude Code’s command deletes is known only when it runs." with Don’t Run Command, Stop Claude Code and Run Command, instead of saying it deletes the folder "/" outside this conversation’s folders.
  • Rogue Agent Protection’s notification, reminder note, channel reply and transcript say the same about a delete whose folder is known only when it runs: for rm -rf $(dirname $d), they now read "Claude Code tried to delete something its command names only when it runs" instead of "outside this conversation’s folders", and no delete question names the folder "/" for it or offers to allow "/" once.
  • After Don’t Allow on a command or a write outside the folders, the model is told it did not happen: when you choose Don’t Allow on the folder question for a command whose file Kyberna can’t know until it runs, or for a write to a file outside the conversation’s folders, the model now reads that the command did not run or the file was not written, and to stop and ask you; before, it read only that you declined.
  • A read-only awk, sed or grep command no longer asks about the folder "/": a slash in a program’s own text, such as an awk pattern or python3 -c code, is not read as a path. When a command builds its path from something known only when it runs, such as "$(git rev-parse --git-path info/exclude)", the question says so instead of naming "/". In a project marked for self build with Self build mode on, such a command runs inside Kyberna’s sandbox without asking.
  • A command whose file Kyberna can’t know until it runs no longer offers Allow This File: when a command writes to a path built from something known only when it runs, such as $L/x.log with L set from another command’s output, the folder question offers Don’t Allow alone and keeps no answer for that file, so the same command asks again next time instead of running with no question.
  • An Anthropic API key not limited to 1 workspace answers: when your organization has 1 workspace, Kyberna finds it and sends it with every request, with nothing for you to do; otherwise the conversation shows 1 line, "This API key is not limited to 1 workspace", saying to create a key for 1 workspace in the Claude Console and paste it in Settings › Models › Anthropic, with the API’s own text under What the Anthropic API reported. Before, every turn failed with the API’s raw "HTTP 400 invalid_request_error" text in a red banner.
  • A local model has 1 name on every sheet that asks you: the approval sheet and the folder question now name it as the held-step sheet does, for example "Qwen3.5 35B wants to run a command"; before, they read "Qwen3.5-35B-A3B-4bit (local)" and "Local: Qwen3.5-35B-A3B-4bit", and a model server’s folder question began with the server’s kind.
  • A model on the Anthropic API has 1 name on every sheet that asks you: the approval sheet and the folder question now name it as the held-step sheet does, for example "Claude Sonnet 5 wants to run a command"; before, they read "claude-sonnet-5 (Anthropic API)" and "Anthropic API".
  • Opening a web page at an address counts as sending for Rogue Agent Protection: when an agent reads a file outside the conversation’s folders and then opens a page in Kyberna’s browser, or a new tab, at an address, Kyberna is 95% sure it went off track and holds its next step, as it does for a web fetch to the same address; in a turn already flagged, that page asks first even when you allowed the browser. Before, a page opened at an address was not counted.
  • Pause and Kill act on a reply Kyberna starts by itself: when a reply ends in an error while your next message waits, and Kyberna goes on to that message by itself, the conversation shows Stop and its spinner while it answers, Pause holds that answer and Kill ends it naming the kill, as for a message you sent; before, the conversation read as finished, Pause said the turn had already ended while the model kept writing, and Kill wrote no line. The same holds when an ACP agent goes on to a message another window sent.
  • A tool your rule denies is refused without a question: on Kyberna’s own engines no approval sheet appears for it, and the refusal names your rule, for example "Settings › Tool Policy is set to Deny for shell_run"; before, a sheet asked first, and the refusal named the profile instead of your rule.
  • The status line under a conversation still says it waits for you when an agent’s question outlasts the reply: it reads "Waiting for your approval" until you answer; before, it went with the reply’s end, so the conversation looked finished while an agent waited.
  • Every audit row about a conversation names its engine and model: kyb gateway audit now shows them on the rows written when a conversation opens and when you send a message; before, those rows showed none until the engine answered, so a turn with no tool call left no audit row naming its engine.
  • A conversation Rogue Agent Protection has paused closes when you close it: the paused action is refused and does not run; before, the close failed and the conversation stayed in the list.
  • A step held in a turn another conversation’s message started names that conversation again: the sheet says "A message from “<title>” started this turn." under the reason it was held; before, that line showed only when no step was held.
  • A held step in a turn another conversation’s message started quotes that message: when the agent gave no reason, the sheet and the delete question read "Claude Code didn’t say why. The message said:" over the message’s own words; before, they read "You asked:", as if you had typed it, over the notice Kyberna adds for the model.
  • A paused run’s notification shows its whole name: a long reminder, task or channel name is no longer cut off mid-word after 60 characters; macOS shortens it only when the banner has no room.
  • VoiceOver reads how sure Kyberna is as text: on a held step’s sheet, the whole-turn sheet and a paused run’s card, the line "Kyberna is 80% sure Claude Code went off track" is read as 1 line of text; before, it was an element of unknown kind.
  • A held step’s first line no longer repeats its subject: when Rogue Agent Protection holds a command, the sheet reads "curl connects to another computer."; before, it read "It curl connects to another computer.".
  • The same prompt sent twice keeps both rows: on every engine, a message with the same text as the one before keeps its row in the conversation; before, the second row was dropped. On Claude Code, 2 messages sent during one turn keep exactly 1 row each, also when the second is sent before Claude Code echoes the first; before, 3 messages could show 5 rows.
  • A conversation the window is not holding open shows the engine it runs on: after Kyberna moves a conversation to another engine, its sidebar row, the Engine filter count and the account line at the foot name the new engine; before, the row kept the engine of its last run in this window.
  • A message queued during a turn that ends in an error is the next message: on Kyberna’s own engines (the Anthropic API, local models and model servers), the engine starts a turn with it, and the model answers it; before, it waited for the next turn’s first tool results or that turn’s end. With a turn limit, a message typed while the model writes an answer that ends on the last permitted model call, or sent during a tool call on that call, is answered as the next turn with its own model-call count; before, the next pass ended "Reached N model calls in one turn" with it unanswered.
  • A conversation waiting behind another request on Kyberna’s model server is no longer cut: the server sends a keep-alive every 15 seconds while the model has produced nothing, so the conversation ends with its answer; before, the engine’s 300-second watchdog, counted from the request, could cut the conversation while it waited.
  • Kyberna’s tools stay on for a conversation after Kyberna restarts: a turn started by a message from another conversation has the tools you turned on, such as Mail, on every engine; before, it had the engine’s default tools and the model’s mail_search call read "Unknown tool".
  • A marker written into a title or a mail header cannot pass for Kyberna’s own: a note’s title, a mail’s From, Subject, Date and attachment names, a web page’s title and address, and a conversation’s title reach the model with any untrusted-content marker neutralized, and the forged text is still audited.
  • Reading text Kyberna fenced is not a forgery, and a forged marker still is: a read of a read no longer sets the boundary-forgery flag; a complete forged pair of markers in a command’s output or a file, or a forged marker in a title or header, now sets it.
  • A conversation that moves into this window runs on the engine it was opened with: before, it could run on the engine the move’s recipe named, for example Claude Code resuming a model-server conversation’s session.
  • Pause while a local model is still reasoning, then Resume, continues its reasoning: on Qwen3.6, a response paused before any answer text was asked again from the start at Resume in 3 of 3 reruns; it now continues from the reasoning written so far, on MLX and llama.cpp. Core AI still asks again.
  • The Claude 5 models on an Anthropic API key show their reasoning: a turn on Claude Opus, Sonnet or Haiku 5.5, Fable or Mythos shows "Thought for N s" with a summary of the model’s reasoning; before, the row did not appear, since the API returned the reasoning empty unless asked for a summary. Claude 4 models answer as before.
  • Claude Code shows its reasoning in every conversation: a Claude Code conversation through Kyberna Gateway shows "Thought for N s" with a summary of the model’s reasoning, as a conversation in the Console’s direct mode did; before, the row did not appear.
  • Claude Code’s question about a website closes when the command stops waiting: when a command Claude Code runs ends before you answer whether it may connect to a host, the question closes at the end of the reply, the conversation says "The command stopped waiting before you answered, so it did not connect to example.com. Kyberna asks again if a command tries example.com again.", and kyb gateway audit names the host as unanswered, as for the other agents; an answer sent after that does nothing. A question from a command Claude Code runs in the background stays open. Before, the question stayed open after the command had ended.
  • A conversation’s title and name are encrypted with the rest of the record: the first thing you type, which Kyberna shows as the conversation’s title, and any name you give a conversation are encrypted on your Mac like its messages, so a copy of Kyberna’s conversations taken off your Mac shows neither. The sidebar and its search show them as before. On a Mac that can’t open the record, the title reads "This Mac can’t open or keep your conversations until you install the recovery key from your other Mac."
  • The guide says what a Codex conversation shows for a file outside its folders: Codex reads the file and lists its folder without asking you, and the conversation shows the command and what it printed; before, the guide said the conversation shows that Codex’s sandbox refused it.
  • A channel’s web fetch of a site nobody named waits for the operator: with Claude Code as the engine, a fetch of a site that nobody writing to the channel named, and that no web search in the conversation returned, waits for the operator’s answer and fetches nothing until then; before, it ran with no question.
  • A reminder’s, a scheduled task’s and a helper agent’s web fetch follows their browser: with Claude Code as the engine, a fetch of a site nobody named runs without asking, as their browser opens such a site, and the audit log lists the site with who fetched it; before, it ran with no audit row.
  • A command whose path Kyberna can’t know until it runs can be allowed once: on Kyberna’s own engines the folder question offers Don’t Allow and Allow Once, and says what the command can reach. Allow Once runs that one command inside the folders and files the conversation can use and temporary folders, so a write or read anywhere else is refused, and it keeps no answer; before, the only answer was Don’t Allow. A later command in the same reply that reaches outside the folders asks again.
  • A conversation whose folder was moved or deleted stays in the sidebar: it shows its whole transcript, read-only, before you choose anything, and choosing another folder keeps it under its project with its transcript and continues it there. Before, choosing a folder dropped the conversation from the list.
  • Settings › Models shows a sign-in finished elsewhere without a restart: the rows of GitHub Copilot CLI, Codex, Grok Build and Antigravity CLI read their sign-in again when you return to Kyberna and when a Settings window becomes key, so a sign-in finished or removed in Terminal or a browser shows at once, and Kyberna offers to turn the agent on once it is signed in.
  • A stored conversation on an agent this version does not offer says so in the sidebar: its account line reads the agent’s name with “Not available in this version” under it, and the tooltip carries the rest, in place of “Sign-in state unknown”.
  • During a meeting, your line no longer shows the others’ words: the microphone’s tentative words that repeat what the meeting app hears are not shown under You, and ones already shown that turn out to be an echo are taken away. Your own words still show.
  • A folder in your home folder reads as ~/… the same way everywhere: the approval sheet’s first line and its wording, the folder question and the Folder Snapshots rows write it alike, including when your home folder is reached through a symbolic link.
  • With Claude Code’s Permission Boundary off, a read outside the conversation’s folders asks the way Claude Code does alone: a call that reads outside the folders, by a read tool or by a command such as ls or grep, is no longer answered for you, so Claude Code asks with its own question at every Approvals level.

0.3.1

October 14, 2026
Download for Mac Kyberna-0.3.1.dmg · 108 MB SHA-256 a30639cf65fe69f27fc5e921bb07787e269f87949bcdb6ee0a4b490bc737fe89

Meeting notes, a move file that carries your conversations to a new Mac, a shorter setup, and mail that goes only to addresses you gave. First launch: open the disk image, drag Kyberna into Applications and open it, confirming once when macOS asks; the setup sheet installs the gateway itself.

New

  • Meeting notes: press Start in File › Start Meeting Notes (⇧⌘M), in the menu-bar item or in the offer Kyberna makes when Zoom, Microsoft Teams or Google Meet starts a call, and Kyberna writes the call down on this Mac, your lines as You and the others named from the meeting’s calendar event. When the call ends, the model you chose writes a summary, the decisions and the action items. No audio is kept, a meeting is kept in Kyberna’s encrypted record or Private, and the first Start tells you the others in the call are not told.
  • One review sheet shows each action item as the reminder, calendar event, note or email draft it becomes, with its text, owner and date editable. The items you tick are added exactly as shown, with Undo, and an email is only ever a draft you send from Mail.
  • Move your conversations to a new Mac: Make Move File… in Settings › General writes 1 encrypted file of every conversation on the old Mac, Kyberna’s and Claude Code’s. On the new Mac, Choose File… and the recovery key you saved on the old Mac add them beside the new Mac’s own, with their projects, reminder records and scheduled tasks (which arrive paused), skipping any already there.
  • Setup is 6 pages with 1 choice each: Welcome, the CLIs and local models to download and sign in to, the default model, where your work goes, your recovery key and a summary. Downloads carry on after you close the sheet, with their progress in the main window, and Run setup again… in Settings › General opens it again.
  • Show Recovery Key… asks for your Mac’s login and then shows the key to copy for this Mac only or to print. Kyberna writes it to no file. If some conversations were recorded under an earlier recovery key, Settings › General lists them under Conversations This Mac Can’t Open, and Enter Key… opens them with that key.
  • Each tool call is its own row in the transcript: it is titled with what the call does, opens to its command or file with a Copy link, shows a check and its duration when it finishes (a red cross if it failed) and a counting spinner while it runs, and a status line under the last row gives the turn’s time, tokens and running calls.
  • A local model, a model server or the Anthropic API can run commands in the conversation’s folders, can run them in the background, and can read a web page you give it. Each command asks first in the Kyberna window and runs in a sandbox that reaches only the conversation’s folders and, through Kyberna’s gate, the network. A page read asks first too, naming the website, reads up to 20,000 characters of its words and does not follow a page that moved to another website. Searching the web is not available yet. A button in the conversation’s header counts the background commands and opens a list where you read, pause, resume or stop each one (⌥⌘P).
  • You see and steer every agent your model starts: a list above the message field shows each one’s state, name, time and what it is doing, with Stop on a running one and Review… on a held one. An agent’s approval request comes up in the conversation that started it, naming the agent and its task, and an agent’s answer reaches the conversation that started it.
  • Your conversations can send each other messages: ask in plain words (“tell the checkout page conversation the format changed”) and a conversation on any engine can list, search or read your other conversations and send one of them a message. The sender’s transcript shows where it went, the receiver’s marks it as from another conversation, and an access key or private key in what one conversation reads of another reaches the model as a placeholder.
  • Undo what an agent did to your files, whatever tool it used: before a turn first changes files, Kyberna takes a snapshot of the conversation’s folder, and Restore… on the turn puts the files back, with Undo Restore afterward. Snapshots are kept 24 hours unless you choose another period or a folder in Settings › Kyberna Guard, and a folder that would take more than 5 seconds asks first. The Rewind button beside the message field is gone, since Restore… does the same on every engine.
  • The browser works out of sight with a tab for each conversation, so a second conversation’s page no longer replaces the first’s. When a page needs you (a password field, a cookie banner Kyberna does not answer, an upload), the conversation waits up to 100 seconds and its transcript row has Show Browser; Take Over holds the conversation’s next call until you choose Hand Back. Helper agents, reminders and scheduled tasks use the browser in a tab of their own while Kyberna is open, and each site they open is written to the audit log.
  • Other apps on this Mac can use Kyberna’s local models: turn on Share Kyberna’s Local Models in Settings › Models › Local models and add each app with Add App…. Each app gets its own key, and Remove… stops 1 app alone. The row says which app is using a model and how many requests each made today. Serve on the Local Network, off by default, lets a device you add reach the models with its own key over an encrypted connection; a run between 2 Macs on a real network is not yet checked.
  • Settings › Developer › Permission Boundary, with Developer mode on, has a row for Claude Code and for each agent you turned on, with the choices Default (Off), On and Off. A change applies to conversations started after it. In Terminal, kyb gateway boundary lists and changes the same choices.
  • Your messages can show in your accent color: Settings › Appearance › Look offers Accent, with your messages in your accent color, and Classic, the earlier gray look.
  • When you ask for a weekly update, a meeting summary, an email reply, a memo, an FAQ answer or a status note, the model can follow a short guide for that kind of writing: the outcome first, a fixed shape, and facts taken from your own mail, notes, calendar and reminders.

Improvements

  • Mail goes only to addresses that came from you: one you wrote, one on the Contacts card of someone you named, one on the message you are replying to or one on a calendar invitation. Any other address waits for you in the Console with a sentence saying where it came from, and in a reminder it is saved as a draft. Approving a send allows that one message, a name that matches several people in Contacts is asked about and never guessed, and a message to more than 20 recipients becomes a draft.
  • Kyberna checks what goes with a message before it leaves: an attached file must be inside the conversation’s folders and must not hold passwords or keys or be a program, and a web address in the message must come from your words or a web search in the conversation. A file that fails stops the message with one sentence naming the file and the reason.
  • When Rogue Agent Protection pauses a reminder, a scheduled task or a channel run, Kyberna Gateway sends a notification with Stop This Run and Let It Continue, even with the Kyberna window closed, and the menu-bar item lists each paused run. A reminder gets a “Paused” block in its notes that you answer with a code, and a channel’s sender is told in plain words.
  • Rogue Agent Protection’s messages say what is held: a download, a build or an in-place file edit says so on the sheet, the notification and the menu-bar item, with buttons named for it, and a held agent’s sheet names who sent the message that started the turn. Rogue Agent Protection also holds more deletes: git rm -r, git clean -fd and srm -r on a folder outside the conversation’s folders.
  • When a conversation and its agents have several questions open, they come one after another, oldest first, instead of being dropped, and each sheet says how many more wait.
  • A conversation that runs in the Kyberna window is held to the same folder rules and Rogue Agent Protection as one on the gateway. Kyberna › Restart Gateway restarts the gateway, and a new version of Kyberna installs and restarts its own gateway the first time you open it, interrupting any conversation running at that moment.
  • Channel tokens, API keys, MCP credentials and the keys you gave other apps are kept encrypted with a key that never leaves this Mac, and the recovery key you copy is cleared from the pasteboard after 60 seconds.
  • For command-line agents, Claude Code, GitHub Copilot CLI, Gemini CLI, Codex, Grok Build and Antigravity CLI decide about files outside the conversation’s folders by their own rules in a Console conversation, and Kyberna asks no folder question for them. Keys, passwords and another app’s private data stay refused, and turning a row on in Settings › Developer › Permission Boundary brings the folder rules back for that agent. When Gemini CLI or Grok Build refuses a file outside the folders, Kyberna asks with Allow This File and Add the Folder.
  • The model picker lists the models of GitHub Copilot CLI and Grok Build under their rows, and the agent runs on the one you choose. Codex’s commands read and write only the conversation’s folders and the files you allowed, like every other engine, and Codex signs in without a copy of its own on the Mac.
  • A message typed while a conversation is paused waits and goes out when you resume; at Kill the waiting messages go back into the message box unsent.
  • Approval sheets name what is asked for every tool, in plain words with a risk grade: a read by an agent names its file, every command, file and web tool of a local, direct or server model has its own sentence, and a reply to a reminder or task opens with the outcome or the decision you need to make. Allow for This Conversation remembers your answer for that conversation, and Always allow lives only in Settings › Tool Policy, where each remembered answer can be removed.
  • Qwen Code and Mistral Vibe are not offered in this version; a conversation you had on either stays in the sidebar and can be read. The 5h and 7d Claude usage rings and their banners are gone until their figures match Claude’s own usage page.
  • A model that repeats itself is stopped, and a reply cut at the output limit says so and is continued up to 3 times.
  • Reminders in a list shared with other people can be marked in Settings › Reminders, so Kyberna warns that the notes may not be yours and drafts, rather than sends, to an address found only there.
  • The memory sheet fits its window and offers only the buttons that help, and when nothing can make a model fit, no button is the default. A conversation, task or reminder you can reply to says in plain words what is waiting and what each line does.
  • Kyberna Gateway carries Kyberna’s icon, so the notifications of reminders, tasks and held runs show it, and a project’s header in the sidebar folds like a folder group.

Fixed

  • Closing an agent while it asks a question no longer hangs, and a plain Stop on an agent’s hold marks the agent as stopped, not failed. An agent’s hold stays on screen when its conversation’s reply ends.
  • Stop works while a question waits for you: Stop, the interrupt and Kill withdraw every open question, so the turn ends and the sheet closes on every engine, and a stop sent to a paused conversation no longer holds back the Resume that follows it. Kill on an Antigravity CLI conversation starts nothing after it.
  • A turn no longer slows down as you keep more conversations: with 4,003 conversations one step took 7.8 seconds and now takes under 0.5.
  • The folder question asks about the right places: it asks about a path a command really touches, never offers Add the Folder for / or a folder that contains your home folder, and says “could not read” for a command that only reads. A file you allow while Antigravity CLI is working is read, and a file in iCloud Drive is named with its own folder.
  • Changing the engine no longer splits a conversation or renames it: every engine’s part stays in 1 history, in order, and the new engine knows the whole conversation so far.
  • GPT-OSS models on this Mac show their reasoning and answer without channel markers and can read files and use tools. A local model that stops before answering says what to do, and choosing another local model or model server in the picker now sends your next message to the one you chose.
  • A message to a conversation that cannot take it reads Not received with the reason, instead of Received.
  • A Private conversation leaves nothing on the gateway after you close it or quit Kyberna.
  • A channel reply carries no image and no link to a site nobody in the conversation mentioned, so a preview cannot send part of the conversation off without a click. A web page can no longer forge what Kyberna reads from it or move a click Kyberna has judged onto another control.
  • Restoring a recovery key either moves all your conversations to it or changes nothing and says why, and the Kyberna window reads the restored conversations at once. On a Mac that cannot open its conversations until you install the recovery key, a message you send meanwhile goes back to the message field instead of being lost.
  • Continue in Claude Code works for conversations with Gemini CLI, Copilot, Codex and other external agents, and comes back with Take Back.
  • File › New Conversation opens the conversation in the window you are in and asks for a folder when you chose to be asked, and a model chosen before the first message is the one the conversation runs on.
  • Don’t Allow and Allow on a helper agent’s email do what they say, and a held channel message always tells both the sender and the operator.
  • Paths with tildes, such as an iCloud Drive path, are no longer drawn as struck-through text, and a long word on an approval sheet no longer breaks with a hyphen that is not in it.
  • Kill ends the detached commands of an agent that can resume a conversation, and Pause and Kill reach a Mail or Notes action that is running. Unloading an MLX model gives its graphics memory back.
  • Git works in a command a local model, a model server or the Anthropic API runs while you have a global git settings file, and its commits carry your name and email. Local engines can compile with swiftc and swift build.
  • In Kyberna Write, Ask Kyberna follows a document to its new folder and name after Save As.

0.3.0

September 29, 2026
Download for Mac Kyberna-0.3.0.zip · 65 MB SHA-256 0097c72df161054a9f5eeb6c0627a121a46425de0d2ce6ad754abf8c9a8fd8b4

Every engine now works inside the conversation’s folders, the command-line agents reach only their makers’ sites and the package sites unless you allow more, and more of what engines send is switched off for conversations in Kyberna’s record or Private. First launch: unzip, drag Kyberna into Applications, then System Settings › Privacy & Security › Open Anyway, once; the setup sheet installs the gateway itself.

New

  • Folder Boundary: every engine, Claude Code included, reads and changes files only in the conversation’s folder and its project’s folders. In the Kyberna window, a read that reaches anywhere else waits for your answer in a sheet: allow that read only, or add the folder. A conversation outside a project keeps the folders you add to it, and the folder pill removes one.
  • GitHub Copilot CLI, Gemini CLI, Mistral Vibe, Grok Build and Qwen Code run behind Kyberna’s own network gate: they reach their maker’s sites and the package sites, any other website asks in the same sheet as Claude Code’s, and the hosts an agent uses only for usage statistics are refused without asking. These agents start only while the Kyberna Gateway runs.
  • Reminders can read files: a reminder reads its own folder, email attachments included, and the folders you list under Folders reminders may read. A read anywhere else waits for your OK, given on the Mac (once, for that reminder, or always) or, in lists you name, by replying with a code in the reminder.
  • Rename a project in place, from its project pill or with Rename Project… on its sidebar heading.
  • Engines update with Kyberna: at the first launch after an update, the newly tested versions of the engines you have on download in the background with 1 notification, and the previous copy runs until they are ready.
  • Grok Build’s row in Settings has Let xAI keep my coding data, which changes that setting on your xAI account after you confirm.
  • Settings › Appearance sets the size of the prompt field’s microphone, send and Stop symbols: 14, 16, 18 or 20 points.

Improvements

  • More is switched off for conversations in Kyberna’s record or Private: Claude Code’s Remote Control, organization memory and session survey; Codex’s analytics and server-side compaction; GitHub Copilot CLI’s remote export and memory; Gemini CLI’s Google web search; Qwen Code’s request metadata and cache; Mistral Vibe’s telemetry; and Grok Build’s telemetry and trace upload, with its sessions kept on this Mac. The conversation’s record note says what is off.
  • An engine reads Ready, and its switch or setup-sheet checkbox turns it on, only once its tested copy is downloaded and verified; the switch downloads first. Antigravity CLI now downloads on any Mac, with its progress shown like the other agents’.
  • The Pinned Claude Code row, and the setup sheet, say when a newer Claude Code for Kyberna is available, with Download, and name the copy that runs meanwhile.
  • Gemini CLI’s models are listed under it in the model picker; when Google is busy, 1 note placed with the turn says how many times it retried and why; a lapsed Google login offers Sign In instead of opening a browser.
  • The Claude usage rings agree with Claude Code’s own usage figures: they show the newest reading from a live turn, with the time it came, and dim a reading over 1 hour old. A turn’s cost is marked “reported by the engine”.
  • Voice: the microphone says Dictating or Voice mode while it listens, “Kyberna, send” and “Kyberna, cancel” work during plain dictation, an edit you make while it listens is kept, and “Hi Kyberna” wakes it like “Hey Kyberna”.
  • A note a conversation reads reaches the model marked as someone else’s words, as mail already did, and hidden characters are removed from a note’s title and a message’s sender.
  • Importing MCP servers asks where the file came from, and each server’s row shows your answer.
  • The sidebar says “Loading your conversations…” while it fills at launch.

Fixed

  • At first launch the setup sheet installs the gateway by itself again, instead of showing an Install button.
  • On a new Mac, the agents that need Node.js (Gemini CLI, GitHub Copilot CLI, Codex, Qwen Code and Grok Build) download from the setup sheet and Settings; before, only Claude Code, Antigravity CLI and Mistral Vibe did.
  • Settings › Models shows the engines you turned on in the setup sheet.
  • Opening Settings › Models no longer holds up the window for seconds.
  • Gemini CLI applies the settings Kyberna writes for it: usage statistics off, your retention period and checkpointing off.
  • An expired Claude Code login shows 1 banner, which goes away once you sign in, and the turn ends as an error rather than a success; an agent that stops at start while signed in no longer shows sign-in steps.
  • Clicking a folder group in the sidebar opens or folds it instead of showing the start page, and a conversation started from the start page keeps the model and record mode it showed.
  • A follow-up you write in a reminder while its answer is still being written or saved gets its own answer, and a reminder being answered when the gateway restarts is stopped and can run again.
  • Nothing keeps animating while a conversation waits for you.

0.2.15

September 28, 2026
Download for Mac Kyberna-0.2.15.zip · 63 MB SHA-256 6a2907f9d0c2588d747c1f0bb950ff54bb437d84f5b80ef53e464d3b18f0f056

A start page when no conversation is open, and Claude Code sending less from conversations in Kyberna’s record or Private. First launch: unzip, drag Kyberna into Applications, then System Settings › Privacy & Security › Open Anyway, once; the setup sheet installs the gateway itself.

New

  • A start page when no conversation is open: a greeting, a box to type in that starts a conversation on your default model and folder (with the microphone when Dictation is on), the model, folder and record mode it will use, a Continue row for your most recent conversation, and up to 4 starting points built from what your Mac has set up.

Improvements

  • Claude Code’s telemetry, error reports and feedback uploads are turned off for conversations kept in Kyberna’s record or Private.
  • New Conversation… and New Task… in the sidebar no longer look disabled.

Fixed

  • A conversation closed and opened again from elsewhere shows the new conversation, not the old one’s messages.
  • File › New Conversation and New Task… open the main window when only the browser window, or no window, is open.
  • A Claude Code copy Kyberna has verified no longer shows a yellow version warning when its version cannot be read at start.

0.2.14

September 27, 2026
Download for Mac Kyberna-0.2.14.zip · 62 MB SHA-256 9a5b965951931bedec936612dd00072fe9adb2cfdf0edd7f788c2e5070bfc40d

A setup sheet that downloads the engines and local models you choose, scheduled tasks, voice, a fuller built-in browser, and the whole app redrawn in macOS 27's Liquid Glass. Kyberna Write now has its own download; this archive holds Kyberna alone. First launch: unzip, drag Kyberna into Applications, then System Settings › Privacy & Security › Open Anyway, once; the setup sheet installs the gateway itself.

New

  • The setup sheet chooses your engines: every engine Kyberna can run is listed with a checkbox, its download size and the terms it runs under, and checking one turns it on and downloads it, adding Node.js where needed. It also downloads the Claude Code version Kyberna was tested with (about 211 MB), offers an open-weight local model that fits your Mac, and installs the gateway the first time Kyberna opens. Every download is checked before it runs, and one that does not match is deleted.
  • Scheduled tasks: a prompt Kyberna runs once, every so many minutes or hours, daily, on weekdays or weekly, each run its own conversation with a result you can read afterwards.
  • Voice: dictate into the composer and have answers read aloud, both on this Mac with nothing sent anywhere, plus a hands-free voice mode.
  • More engines: Google's models 3 ways (the Gemini API, Gemini CLI and Antigravity CLI), Codex in its "Ask for approval" mode with its own model picker, Mistral Vibe signing in with an API key, and Mistral Vibe, Grok Build and Qwen Code downloaded and verified like the other engines. Core AI models run on Apple's own on-device runtime as a third kind of local model.
  • The built-in browser gains a private window (⇧⌘N) that keeps nothing once it closes, file downloads to your Downloads folder with a downloads list, page dialogs shown as sheets, Safari's keyboard shortcuts, and a list of safe websites a conversation can open without asking. It never accepts cookies for you: a conversation declines what a site does not need, or leaves every banner to you.
  • Bring your own MCP servers to Claude Code conversations, imported from a Claude Desktop, Cursor or Cline configuration file. Each arrives off; turning one on pins its exact version, lists its tools and holds back any tool description that reads like an instruction to the model.
  • Pause, Resume and Kill reach everything a conversation runs, its commands and background agents included, and a new menu-bar item pauses or stops everything Kyberna is doing, reminders, tasks and channels included, even with the Kyberna window closed.
  • Keep one conversation longer or shorter than the default, from its record pill; the result line under each turn gives its cost when billed per token, with the engine, provider and model in its tooltip.

Improvements

  • The whole app is redrawn in Liquid Glass for macOS 27, with smaller pills at Safari's size and macOS's standard sidebar button in every window.
  • A cleaner sidebar: a New Conversation… row at the top, one Engine filter, a conversation shown by its title alone, groups that fold with macOS's own disclosure, and an order that holds still until you quit. The list opens in 104 ms with 10,000 conversations, down from 5.2 seconds.
  • Messages you send while the model works wait as bubbles at the foot of the conversation and reach it as soon as the running step finishes, with Send now on each.
  • Approval requests in plain words ("Claude Code wants to run git status in ~/Kyberna"), colored by how much the call could change, with the raw call behind Show details. Fewer of them: one answer covers a site, a kind of command or a command prefix such as "git push", and installing a package from the Python, npm or Homebrew sites no longer asks. Sending, network and browser calls always keep at least the risky-call question.
  • The browser reads a whole list that loads as you scroll, such as a feed or a team page, in one action, and the conversation window comes back in front with the keyboard in the composer when a browser turn ends.
  • A turn that goes quiet is retried and resumed on its own, and a stuck turn is stopped after 30 minutes.
  • When a local model does not fit, the memory sheet offers to unload Kyberna's own idle models first, and 2 conversations starting large models together can no longer both pass against the same free memory.
  • Reminders are checked a few seconds after the Mac wakes, and a completed reminder's conversation follows your retention period like any other.
  • Claude Code 2.1.281. Shift-Return inserts a new line in the composer. About Kyberna shows the build, the release date and a link to these notes.

Fixed

  • A gateway restart no longer asks you to start conversations again: each reconnects on its own and carries on, and an open approval stays answerable.
  • Installing a package no longer hangs on a question you never saw: the sheet names the website the command wants to reach.
  • Allow Calendar… and Allow Contacts… grant access again, and a contacts lookup no longer brings the gateway down.
  • The gateway starts while the screen is locked, survives a program that exits mid-write, and reports a socket path that is too long instead of crashing.
  • A shell command in a conversation run without the gateway now asks before it runs, and the credentials of the browser and your MCP servers are no longer visible in the process list.
  • After driving the browser, the keyboard returns to your window; the engine picker scrolls when the list is long; the microphone button stops dictation at once.
  • Reminders are answered on their own again without Scan Now, and a conversation in Kyberna's record opens at its end.

0.2.13

September 21, 2026
Download for Mac Kyberna-0.2.13.zip · 63 MB SHA-256 94eb03468011060639e349ccec0646e95358aa18e7f0d1650ab2c345a513c43a

A redesigned Settings window, recovery of interrupted model downloads, and an API-key route for Gemini CLI. First launch is unchanged: unzip, drag both apps into Applications, click Done on the dialog, then System Settings › Privacy & Security › Open Anyway, once per app.

New

  • Settings is a two-pane window: a sidebar with General, Appearance, Engines & Models, Tool Policy, Channels, Reminders and Session Sync under "Settings", Developer under "Advanced", a search field that narrows the sidebar to the sections holding a matching row, and a one-line description under every row.
  • Model Catalogue downloads that finished while the app was closed are picked up at the next launch, and "Recover Downloaded Files" in the Downloads pane (also kyb models recover) matches finished files left in the system's download cache against open downloads by size and checksum and moves them into place, deleting nothing.
  • Gemini CLI takes a Gemini API key entered in Settings › Engines & Models › Agents (or kyb agents key gemini), stored beside the Anthropic key and passed to that agent alone. Google closed the personal Google-account sign-in for Gemini CLI on June 18, 2026, so the key is the remaining route. GitHub Copilot CLI 1.0.86 ships pinned like Gemini CLI.
  • A tool can carry its own risk judgment of a call (the SDK's riskOfCall, KybernaAgentKit 0.7.0); the file and mail tools do, and the approval sheet reads it first.

0.2.12

September 21, 2026
Download for Mac Kyberna-0.2.12.zip · 63 MB SHA-256 032c30477c2cc45d7c01694bd4a7c0059218010aa108776454e596a06c6ff7ad

Tool policy with patterns and risk grades, read-only file tools and project instructions for local models and API keys, conversation compaction, and a first engine over the Agent Client Protocol. First launch is unchanged: unzip, drag both apps into Applications, click Done on the dialog, then System Settings › Privacy & Security › Open Anyway, once per app.

New

  • Tool policy rules may hold wildcards (mcp__kyberna__mail_*), the last matching rule in a profile wins, and every pending call carries a risk grade (low, medium, high, unknown) with its reasons shown on the approval sheet; shell commands are graded without running them. A profile's confirmation setting can let low and medium calls run on their own while high and unknown still ask. "Always allow" now remembers the tool with its arguments, so the same tool with other arguments asks again.
  • Local models, API keys and LM Studio's or Ollama's servers get read-only file tools (read, glob, grep) confined to the conversation's folder and its project's folders, and read the project's AGENTS.md or CLAUDE.md the way Claude Code does.
  • Conversations on those engines compact themselves at 70 percent of the model's window: the middle is summarised, the newest quarter stays verbatim, and long tool output spills to a file the model can read back. The transcript shows where it happened.
  • Engines › Agents: Gemini CLI, GitHub Copilot CLI, Codex, Mistral Vibe, Grok Build and Qwen Code over the Agent Client Protocol, each off until turned on in Settings, with the sign-in note the vendor documents. Gemini CLI 0.60.0 ships pinned like the Claude Code engine; the others run through npx at an exact version.
  • Window › Model Catalogue has a Loaded Models pane: which local models are in memory, which conversations hold them, and what LM Studio and Ollama hold, with an Unload button. Also kyb models loaded.
  • The folder pill's menu has "Choose another folder…" and "Reveal in Finder"; a conversation without a project relocates in place.
  • Notes: the tools skip Recently Deleted unless asked; the mail tools' descriptions open with "Email" so small models pick the right one.

Fixed

  • The memory check before a local model loads now counts Jan, text-generation-webui, LocalAI and llama.cpp's own server as present without offering to stop them, and reads Ollama's merged vision models as unsupported with the reason instead of failing at load.

0.2.11

September 21, 2026
Download for Mac Kyberna-0.2.11.zip · 62 MB SHA-256 b8046464072e8d36b6c0ea2d0dbe49a1fae6a6ab3b7ed40d5eb14f7e780af963

Projects, a default working folder, and a conversation list that no longer depends on the Claude Code CLI. First launch is unchanged: unzip, drag both apps into Applications, click Done on the dialog, then System Settings › Privacy & Security › Open Anyway, once per app.

New

  • Projects. A project holds several working folders and several conversations. The conversation header shows its title, its project and the current folder; the folder pill switches among the project's folders mid-conversation, and the transcript records where the switch happened. The sidebar groups conversations by project, with a "No project" group for the rest; a conversation without a project can be added to one, once, from its header. Also from the terminal: kyb projects.
  • A default working folder, offered as ~/Kyberna during setup and changeable in Settings › General; new conversations open there instead of a hidden scratch folder.
  • Conversations on a local model, an API key, or LM Studio's and Ollama's servers now appear in the sidebar, can be renamed, keep their names, and reopen with their history, on a Mac with or without the Claude Code CLI installed. Notes: a tool that lists notes changed in the last N days.
  • The result line under a local model's turn shows how much of the prompt came from the cache.

Fixed

  • Clicking a Claude Code conversation that had been switched to a local model showed another conversation's title, added a row to the sidebar each time, and could not be renamed.
  • A small local model that wrote its tool call in an unusual order leaked the raw call text into the answer.

0.2.10

September 20, 2026
Download for Mac Kyberna-0.2.10.zip · 61 MB SHA-256 c607392a4ea0791adb7d9a242a2c823560828a20942f1ef2173021dc6ca604fa

Faster local models and two small fixes, released the same day as 0.2.9. First launch is unchanged: unzip, drag both apps into Applications, click Done on the dialog, then System Settings › Privacy & Security › Open Anyway, once per app.

Improvements

  • A local model is loaded once and shared by every conversation on it, and the part of the prompt that has not changed since the last turn is not read again. Follow-up turns start far sooner; the result line shows the tokens per second.
  • When a conversation has thinking turned off, models of the Qwen family are now told not to reason instead of reasoning out of sight.

Fixed

  • After a few idle minutes a conversation said "The engine finished" and needed Start again. The engine now restarts by itself with your next message.

0.2.9

September 20, 2026
Download for Mac Kyberna-0.2.9.zip · 61 MB SHA-256 659d0dbcd3459a8ad9e5270d768ef6fdf85f3fdeb5218938d66ed8ed58960d20

Live output for local models in the Console, released the same day as 0.2.8. First launch is unchanged: unzip, drag both apps into Applications, click Done on the dialog, then System Settings › Privacy & Security › Open Anyway, once per app.

Fixed

  • A local model's reasoning and answer appeared all at once when the turn ended. Both now stream into the conversation as they are written, the reasoning into its own row.

0.2.8

September 20, 2026
Download for Mac Kyberna-0.2.8.zip · 60 MB SHA-256 4b26a3a54192f98fc2652061bd81ea2fe8a0256c12b5dafa0cfca81cfad819aa

One fix to how a local model's reasoning is shown, released the same day as 0.2.7. First launch is unchanged: unzip, drag both apps into Applications, click Done on the dialog, then System Settings › Privacy & Security › Open Anyway, once per app.

Fixed

  • With some local models (Qwen 3.5 and 3.6 among them) the model's reasoning appeared as part of its answer, ending in a stray closing tag. The reasoning is now separated and shown the way the Claude desktop app shows it: in place, in italic, behind a thin bar, with how long it took.

0.2.7

September 20, 2026
Download for Mac Kyberna-0.2.7.zip · 60 MB SHA-256 f71f27671413e8ffd49d32d84f2c50a9d5c17aff4e8365d39ba06d95de3ec92c

Local models: a way past the memory check, and clearer handling of files Kyberna cannot run. Released the same day as 0.2.6. First launch is unchanged: unzip, drag both apps into Applications, click Done on the dialog, then System Settings › Privacy & Security › Open Anyway, once per app.

New

  • When the memory check says a local model does not fit, the sheet now offers Load anyway. The estimate stays on screen; macOS may compress or swap, and the load may fail if memory truly runs out.

Fixed

  • Some models pulled by Ollama are stored in Ollama's own merged form, with a vision tower Kyberna's runtime does not read; they failed to load with no reason given. They are now listed greyed out with the reason and a pointer to running them through Ollama's server, and any load that still fails reports the runtime's own error.
  • After a local model failed to load, the engine picker lost its checkmark and named the model by its file hash. The picker now keeps the model's name and row.

0.2.6

September 20, 2026
Download for Mac Kyberna-0.2.6.zip · 61 MB SHA-256 c98148c7ec80831a326299045f784fb039d6e7bede187c008e56c8dfe01a910a

One fix to the memory check before a local model loads, released the same day as 0.2.5. First launch is unchanged: unzip, drag both apps into Applications, click Done on the dialog, then System Settings › Privacy & Security › Open Anyway, once per app.

Fixed

  • Models whose files describe attention per layer, Gemma 4 among them, were reported as needing several times their real memory (a 27.5 GB model as 156.6 GB) and refused. The estimate now reads those per-layer figures, and the same model reads 31.7 GB.

0.2.5

September 20, 2026
Download for Mac Kyberna-0.2.5.zip · 61 MB SHA-256 d8cd147e020f73012e26ced37f486231a2de4146a131d353d23caaa4f9ac8825

Download progress in the Model Catalogue and a set of small fixes for the two-Mac tooling, released the same day as 0.2.4. First launch is unchanged: unzip, drag both apps into Applications, click Done on the dialog, then System Settings › Privacy & Security › Open Anyway, once per app.

Improvements

  • The Model Catalogue's model page now shows each file's download progress with its speed, a pause button, and after a failed download (a dropped connection, say) the reason with Resume or Retry. A model made of many files shows one bar over all of them and a row per file.

Fixed

  • Reminder test runs no longer leave their sessions in the sidebar, and the test runner's help option no longer starts a run.
  • Tooling for two Macs: the relay refuses a reply from an older agent instead of reporting a wrong exit code; setup checks for Xcode's Metal Toolchain before building; pinning an already-pinned engine leaves the lock file unchanged.

0.2.4

September 20, 2026
Download for Mac Kyberna-0.2.4.zip · 61 MB SHA-256 41db46cd12ee94aa378b9b516ac81f00727eaa3f3fb9bf95244c98758ad15cd0

A change to how local models get Kyberna's tools, released the same day as 0.2.3. Everything else is as in 0.2.0 below. First launch is unchanged: unzip, drag both apps into Applications, click Done on the dialog, then System Settings › Privacy & Security › Open Anyway, once per app.

New

  • Kyberna's own tools (Mail and Notes) are now a per-conversation switch, the last row of the engine picker. They stay on for Claude models and start off for local models, where their descriptions took up most of a small model's prompt. Turn them on when a local model should read or draft mail and notes.

Fixed

  • A small local model answered "hello" with a list of mail and notes operations. With the tools off by default for local models, and one line per tool when they are on, it greets back.

0.2.3

September 20, 2026
Download for Mac Kyberna-0.2.3.zip · 61 MB SHA-256 fe4ad6ba80b2c10b0ec2575da65731d3f4435bcfdc68f3dd3591d5348de6e937

One fix in the Console, released the same day as 0.2.2. Everything else is as in 0.2.0 below. First launch is unchanged: unzip, drag both apps into Applications, click Done on the dialog, then System Settings › Privacy & Security › Open Anyway, once per app.

Fixed

  • Before loading a local model, Kyberna checks whether it fits in memory. The check counted only free memory and left out the memory macOS keeps as file cache and gives back on demand, so on a Mac with tens of gigabytes unused a small model could be reported as tight or as not fitting. The check now counts that memory too.

0.2.2

September 20, 2026
Download for Mac Kyberna-0.2.2.zip · 61 MB SHA-256 9966bcfd6aaeb739534d18528d69a399c5578d72449053846ef5eb0a6c789946

Two fixes in the Console, released the same day as 0.2.1. Everything else is as in 0.2.0 below. First launch is unchanged: unzip, drag both apps into Applications, click Done on the dialog, then System Settings › Privacy & Security › Open Anyway, once per app.

Fixed

  • In a conversation running in the Gateway, pressing Start again after the engine had stopped showed every answer and every result line twice from then on. Each now appears once.
  • The line under a running turn said "Waiting for Claude…" whatever was answering. It now names the model, for example "Waiting for Qwen3.5 0.8B…" for a local model or the picker's name for a Claude model.

0.2.1

September 20, 2026
Download for Mac Kyberna-0.2.1.zip · 61 MB SHA-256 3b154112f3b96e763eeeb0548a829f5599f7b2e439798f902ba397958dec605c

A fix for 0.2.0, released the same day. Everything else is as in 0.2.0 below. First launch is unchanged: unzip, drag both apps into Applications, click Done on the dialog, then System Settings › Privacy & Security › Open Anyway, once per app.

Fixed

  • Installing the background Gateway from the setup sheet did not work in 0.2.0: the Gateway started and quit at once because the local-model runtime was not copied alongside it. 0.2.1 copies it, and the Gateway starts. If you installed 0.2.0, install 0.2.1 over it and press Install in the setup sheet again.

0.2.0

September 20, 2026
Download for Mac Kyberna-0.2.0.zip · 60 MB SHA-256 bab8fd9079e04737b257d00271d3c1afc6d71896facc3838c4f731e226610926

The archive grew from 13 MB to 60 MB because it now carries the llama.cpp framework and the MLX Metal library needed to run local models.

First launch: unzip and drag Kyberna and Kyberna Write into Applications before opening them. This build is not yet signed with an Apple Developer ID, so macOS says it cannot verify the app: click Done, open System Settings › Privacy & Security, scroll to Security and click Open Anyway next to the app's name (once per app).

New

  • Kyberna can now run without a Claude Code subscription: connect an Anthropic API key, an AWS, Google Cloud or Microsoft Foundry account, or a fully local open-weight model, and switch between them per conversation or per channel.
  • Local models run on your own Mac, using the llama.cpp and MLX runtimes, or point at models you already downloaded through LM Studio or Ollama, including LM Studio's own server.
  • A built-in model catalogue searches, downloads, verifies and manages local models, and checks that a model will fit in memory before it loads, warning you first if it would not.
  • If a model's usage runs out mid-conversation, Kyberna automatically switches to the next model in an order you set, and keeps the conversation going without losing your place.
  • Type /model in the Console or in a channel to see every model and engine you are allowed to use, and switch between them by name.
  • Ask Kyberna: select text in any application, right-click, and choose Ask Kyberna for an answer in a small floating panel, with the option to continue the conversation in the full Console.
  • A guided, built-in browser walks you through connecting Slack step by step, handling every step except sign-in and your own tokens.
  • Paste a chat conversation copied from another app (⌘V in the sidebar) to start a new Kyberna conversation from it.
  • Channels such as Slack can be set to one of three modes — read-only, assist or full — instead of configuring every tool's permission individually.
  • Kyberna Write and the Console now print through the standard macOS print panel, with page numbers on every page and in PDF exports.
  • The Markdown source view now styles headings, bold, italics, code, links and lists as you type, while the saved file stays plain text; the same editor is used for the message composer.

Improvements

  • Reminders testing no longer rings your devices, and each test scenario now runs against a fixed model so results are repeatable.
  • Kyberna's notes and mail tools are safer: a note is only appended if it has not changed since it was last read, and mail is only sent to an address you have allowed or in reply to a thread you started — everything else becomes a draft for you to send.
  • A development build of the Console now keeps its own settings and permissions separate from your everyday installed copy.

Fixed

  • Search now scrolls to and highlights every match, with a running count, in the Console transcript and in both the Source and Rendered views of Kyberna Write; it previously found matches without showing them.
  • The transcript layout options for row spacing, answer shading and corner rounding now visibly change the layout as described; previously moving them had no visible effect.

Security

  • The passphrase behind session sync between your two Macs is now sealed with a key held in your Mac's Secure Enclave rather than kept in a plain file; a copied file is useless on any other Mac.

0.1.1

September 18, 2026
Download for Mac Kyberna-0.1.1.zip · 13 MB SHA-256 8a8b31cf6462b9b124341e94ef53182b9460f59029e0b10b4b7fff1a8f309742

First launch: unzip and drag Kyberna and Kyberna Write into Applications before opening them. This build is not yet signed with an Apple Developer ID, so macOS says it cannot verify the app: click Done, open System Settings › Privacy & Security, scroll to Security and click Open Anyway next to the app's name (once per app).

Security

  • The session-sync settings were hidden from Settings for beta testers, because this unsigned build kept the sync passphrase in a plain file rather than a protected store; they return once that no longer applies.

0.1.0

September 18, 2026

Not available for download: this archive is no longer hosted.

New

  • First public download of Kyberna and Kyberna Write.