Release History

Saved keys now look saved. A small follow-up to 0.6.18, based on how the new key screens felt in use.

• A saved key has its own clear row. Instead of faint text in an empty-looking box, each saved key shows a green Saved · ends in abcd with Replace and Remove. Replace opens an empty box for the new key, and Cancel leaves the saved one as it was. Models you added yourself show their saved key the same way, with Remove key.
• The eye button appears only while you type. It lets you check a key you have just pasted. Saved keys are never shown on screen.
• Your Google plan shows up reliably. The Usage tab could say no Google key was connected when one was. If the first check is missed, it tries again once LLypses is ready and when you open the tab, one check at a time, and it says plainly if Google turns the key down.

Your API keys are now encrypted on your computer, and Settings no longer shows them. Your saved keys are locked to your own Windows account automatically, the first time this version starts. There is nothing to re-enter.

• Your API keys are encrypted. They are protected with Windows' own encryption, so only your Windows account on this computer can read them. A copy of your settings on another computer, or under another account, is useless. If you ever go back to an older version, paste your keys in again.
• Settings never shows a saved key. Each box now says Saved — ends in abcd instead of showing the key, and a new Remove button takes a key out. Models you added yourself work the same way, with Remove key.
• A key that does not save now says so. Settings, the first-run setup and Add provider used to show Saved even when a key was turned down. They now show what went wrong and keep what you typed.
• Member portraits and your logo stay put after a full restart. They could come back blank after closing LLypses completely. Nothing was lost; LLypses now waits until it is ready before loading them, and tries again if it misses.
• Send to Workshop works in every project. The instructions it writes for your coding assistant are now complete on their own, and LLypses only ever changes its own section of those files. An optional Extra workshop instructions file in Settings → Workspace points to your team's fuller instructions.
• Credits for the software LLypses is built on. The licences of the open-source projects inside LLypses are now in Updates & Feedback → Third-party notices.
• Tidier under the hood. We cleaned up the code to our company standards and removed files the app never used. The manual now explains how much active Knowledge Base material fits into a turn.

An important security fix. Please consider replacing your API keys. A flaw could have let a web page you visited, or somebody on the same network, read the API keys saved in LLypses. It is fixed. Nothing about your boardroom, Members, sessions or knowledge base has changed.

• A security flaw that could expose your API keys is fixed. LLypses runs a small service on your computer so the app can talk to itself. Until now it accepted requests from any web page open in any browser, and from any other device on your network — without checking who was asking. A malicious page could have asked for your saved settings and received your provider keys in plain text, read your sessions, browsed folders on your disk, or run board turns billed to your account. It now refuses anything that is not LLypses itself, on both the ordinary connection and the live one carrying your board conversation.
• We suggest replacing your API keys. We found this in an audit of our own code, not in the field, and we have no evidence anyone was affected. But the exposure was real while it lasted. Creating a fresh key at each provider, revoking the old one and pasting the new one into Settings is the only way to be certain. If that is a lot of work, the risk of doing nothing is low — but the choice should be yours to make knowingly.
• LLypses no longer closes other programs on your computer. If another program was using the same connection LLypses needs, starting LLypses closed it — silently, with nothing in any log. LLypses now checks whether it found one of its own leftover processes before closing anything, and leaves anything else completely alone.
• Uninstalling is safe even when it fails. The uninstall screen deleted everything you had not ticked before checking it could actually uninstall. If it could not, your keys, sessions and knowledge base were permanently gone and LLypses was still installed. It now starts the uninstaller first and deletes nothing unless that succeeds.
• Uninstalling now removes Member portraits too. The pictures you upload for Members were covered by none of the tick boxes, so they stayed behind even if you asked for everything to go. They are now included under Generated images, member portraits & company logo, and that option has been renamed to say so.
• Two error messages that were themselves broken. Reporting a problem while offline now shows the message it was always meant to show, instead of failing. The same fault stopped board reports arriving when the knowledge base could not be read; they now arrive without that portion rather than failing outright.

A new Tron theme, a smoother feel throughout, and @ARCH now keeps up with the newest models. A polish release — one new look, motion and depth across the whole app, and fresher model intelligence working behind the scenes.

• New "Tron" theme. A near-black Grid look with an electric cyan accent and a complementary orange frame. Choose it in Settings under App Theme, alongside Synthwave — its swatch is split cyan and orange so it is easy to spot.
• The whole app feels more finished. Board replies ease in as they arrive, buttons and controls answer with a subtle lift, and menus, empty screens and loading states are cleaner. Nothing moved and nothing was renamed. If you have reduce motion turned on in Windows, LLypses honours it and holds still.
• @ARCH keeps up with the newest models. When @ARCH writes a member for you, it now understands the latest OpenAI models — so a member you build on one of them gets a prompt tuned to how that model actually works.
• Auto-detect shows your folder path the normal way. The Workspace folder finder was displaying your projects path with the wrong slashes. It now shows them the usual Windows way; nothing about where your projects live changed.

Give every member a face, and see who is speaking. Members can carry a picture now, and it follows them into the boardroom.

• Members can have a picture. Give any member a portrait on the Roster tab — drop in an image of any size and LLypses scales it down for you. It shows on the roster and now right beside the member as it speaks, and greys out when a member is sitting a round out.
• See who is speaking. The member holding the floor is marked as it happens — a green dot on its picture and a soft glow down the side of its reply — so a fast-moving board is easy to follow.
• A member that cannot see an image now tells you the fix. A member whose provider could not receive an attached picture used to guess "try sending it again", which changes nothing. Now it points you at what works: switch that member to a different model, or — if only the file format was the problem — save the picture as a PNG or JPEG and attach it again.

Maths now reads as maths, and your figures are left exactly as you wrote them. Ask a member anything with maths in it and the raw markup showed through where the symbol belonged. The answer was always right; only the way it was written out was wrong.

• Maths symbols render properly. Members write maths in a notation that wraps each symbol in dollar signs, and LLypses printed that notation out character by character — a heading reading 1. The Mathematical Constant ($\pi$). Symbols now come through as symbols in headings, in bold, and mid-sentence.
• Amounts of money are never touched. The same dollar sign marks a price, and a rule loose enough to catch every symbol is loose enough to reach from the $ in $5M to the $ in $10M and rewrite what lies between. The rule is deliberately narrow, and revenue grew from $5M to $10M comes through exactly as written.
• Anything it does not fully understand is left alone. Fractions, roots and sums with limits are shown as they arrived rather than half-converted, because a half-converted formula reads as though the member said something garbled. Maths inside a code block is left alone too.

Reopening a session now puts you back in the room you held it in. If you ran a session with a room of your own and came back to it later, LLypses seated the full board instead — the conversation was all there, but the people around it were the wrong ones.

• A session reopens in its own room. The same members, the room's name above the list, and the same setting for whether they answer when you address nobody. The room is recorded on every turn, so a session you moved into a different room part-way through reopens in the room you finished in.
• Sessions from before this release say so. There is no honest way to work out which room an older session was held in — a member can sit in a room for an hour and never speak, so guessing from who talked would be wrong often, and a wrong room is worse than an obvious gap. Those sessions reopen with the full board, as they do today, and a line above the member list says their room could not be restored.
• A quiet room no longer silences the next one. A room set not to answer unless addressed used to keep that setting after you left it, so a new session came up with a full board that stayed silent and nothing to explain why. Leaving a room now clears its settings along with its members.

A long answer only existed on your screen until the member finished writing it. If you pressed Stop, or LLypses stopped before the member did, everything written up to that point was gone — including the words you were reading. Replies are now saved as they arrive.

• Pressing Stop no longer throws the answer away. You could watch a member write for three minutes, press Stop because you had what you needed or by accident, and lose every word. What the member had written is now kept, in the transcript and in anything you export, marked underneath as a reply you stopped.
• An answer that gets interrupted now survives. Replies are saved continuously while they arrive, so the rare case where the board goes quiet part-way through an answer costs you a restart instead of the work. Reopening the session shows what the member had written, marked as a reply that never finished.
• The board is told when an answer was cut short. A half-finished reply that looks finished is worse than a missing one — the next member would read a sentence that stopped mid-thought as the speaker's settled position and build on it. Members now treat a cut-short reply as unfinished. Ask the member to carry on from where it stopped and the room follows.
• None of this costs you anything. Saving as it goes sends nothing anywhere, does not use your API key more, and does not change what a turn costs. Every session you already have opens exactly as before.

LLypses was telling you that working models would not answer, and the model list itself had fallen a generation behind. Newer Gemini models were marked "unavailable on your key" in the Roster — models your key reaches perfectly well, and which answer in the Boardroom if you use them. Every model added or removed in this release was tried on a real key first.

• Roll Call now counts a model that answers as reachable, even when our own reply limit cut the answer off before a word appeared. To stay cheap, the check asks each model for a very short reply — and the newest Gemini models think before they answer, so they can spend that whole allowance thinking and finish without a visible word. Nothing failed and nothing was wrong with your key, but the check was looking for words, saw none, and wrote the model off.
• The wrong answers already saved are cleared for you. Roll Call results are kept for up to a week, so a corrected app could still have shown the old answer for days. LLypses now recognises results that came from the version that got this wrong and discards them. It re-checks once on first launch, and there is nothing for you to press.
• You may see a better suggested model than yesterday. LLypses had been steering around models it wrongly believed were closed to you.
• An unreadable refusal is no longer treated as a broken model. If a provider responds in a way LLypses does not recognise, that means we could not read the answer — not that the model is unavailable. The Roster now says nothing rather than guessing. A genuine refusal, where the model does not exist or your key is not entitled to it, is still reported exactly as before.
• Nine models added, each confirmed to answer on a real key. Gemini 3.7 Flash and Gemini 3.6 Flash from Google; GPT-5.6 Sol, Terra and Luna from OpenAI; Grok 4.6; and DeepSeek V4 Pro and V4 Flash. Luna is the one to look at — close to the flagship on capability at roughly a twenty-fifth of the running cost, which makes it the natural home for members on the smaller GPT-5 models. Open the Model Library and press Sync to pick them up.
• Auto had been choosing the weaker model. The ranking behind Auto and Performance mode was assembled by hand and had drifted, so for two generations of models it was confidently handing members the weaker of two options. It is now derived from an independent capability index rather than chosen, and every deliberate departure from it is written down. Auto members may move to a different model after this update. A model you pinned by hand is untouched.
• Five models removed, each one proved unable to answer. GPT-5.5 Pro, GPT-5.4 Pro and GPT-5.3 Codex are served by OpenAI through a different kind of connection than the one LLypses uses, so they could never have answered here on any key. Gemini 2.5 Flash Lite is no longer opened to new users, and Grok 3 on OpenRouter was retired in May. Published documentation claimed ten other models were also dead; all ten answered when tested, and all ten stay. A saved room naming a removed model still opens — the member just needs a new model.
• Your own custom models are no longer cut short. A custom model you added was being capped lower than any built-in one, at a quarter of the allowance, so long replies stopped mid-sentence under a "Cut off here" note with nothing wrong at your end. LLypses now sends no reply limit at all for a custom model — it does not know your server's limits and should not be inventing them.

Pressing Brief on a full board asked every member to write the same summary. Eleven members meant eleven summaries of one meeting, and eleven replies you paid for. Brief now does what it always looked like it did.

• Brief is one summary again. It goes to a single member, once — so it costs one reply whatever the size of your board, and arrives far faster. Nothing about how the board discusses has changed; only the summary at the end.
• The brief keeps the detail. Every retelling of a meeting loses a little, and ten retellings lost enough that one member's correction was undone by another member's summary of it. One brief keeps the numbers and specifics as they were said, and where the room disagreed it says so and says who held which position rather than averaging it into a consensus that never happened.
• Your knowledge base stays out of the brief. A brief summarises the meeting you just had. Sending your reference material along with it was both an expense and a way for material nobody discussed to turn up in the summary.
• Members no longer show you the raw text of a failure. An unfamiliar provider error could put the provider's own internal text in the reply, cut off mid-word. If a provider explains itself in plain language you still get that sentence; anything else is kept back and you get a clear line instead.
• A speech-to-text model is no longer offered as a member's brain. It cannot hold a conversation, and a member assigned to it would have failed on every turn for no visible reason.
• gpt-5.3-codex stops disappearing from your dropdowns. Syncing your Model Library quietly removed a model LLypses ships with, mistaking it for a retired OpenAI family. A member set to it lost its model with nothing explaining why.
• Save is where you can see it. Syncing ticks the new models, but saving is what puts them in your member dropdowns — and the Save button was off screen once you scrolled into the list. A prompt now follows you down the list, says how many are waiting, and saves them.

Three people said they did not know what an API key was, or which AI models LLypses could run. Both answers were supposed to be in the app. One was hiding; the other genuinely was not there. This release is about the first hour after you install.

• A guided tour that walks you through the actual app. Instead of a few screens of text on a grey background describing controls it was covering, the welcome now highlights the real thing, one step at a time, with Next, Back and Skip. And it opens each part of the app as it goes rather than pointing at the tab and describing what is behind it — you land on Roll Call itself, on the Knowledge list itself, on the model picker itself. Fourteen stops, ending back in the Boardroom wherever you leave it. It runs once and then leaves you alone — later updates will not make you sit through it again — and you can replay it whenever you like from Help → Replay welcome tour. If you already own LLypses you will see it once on this update, which is deliberate: it is new, and its last stop is the model list below.
• Setup is one screen now, not three. The only part of the old welcome that actually did anything — connecting your first AI provider — is all that is left of it. Everything else became part of the tour.
• Every model LLypses can run, in one place. A new section in Help lists all ten providers and every model under each, with a mark against the ones your keys already reach. Previously the model list only showed providers you had already paid for a key with, which meant the easiest way to find out what LLypses supported was to buy something first. No key or internet connection is needed to read it.
• "New to API keys?" is a button now. It looked like a notice, so nobody pressed it, and the guide behind it went unread. It says what it does, shows the providers it lists before you open it, and opens by itself if you have not connected anything yet.
• Fixed: syncing your Model Library could not actually give you the new models. When a provider released something new, Settings → Model Library → Sync found it and showed it — but it arrived switched off, so the model never reached the dropdowns on the Roster tab and there was no way to tell that a step was missing. Roll Call and Verify could not help, because neither of them changes your library. New models are now ticked for you the moment a sync finds them, with a line saying how many turned up and that Save is what puts them in front of your members. A sync that fails or times out no longer quietly restores models to a provider you had trimmed. And if you clear a provider down to nothing, LLypses now warns you before saving — that removes it from every member dropdown, even though your key still works.

A patch, and all of it comes from one accident. A project directory full of code libraries got added to a Knowledge Base — twenty-two thousand files of it. 0.6.7 already stops that happening. This is everything else that went wrong on the way back out.

• Clearing a lot of entries is fast now. Deleting twenty-two thousand entries took minutes, with nothing on screen to say whether it was working or stuck, because LLypses saved each one separately. It is one operation now, and the same clear-out takes seconds. Switching many entries on or off, and moving many into a folder, were fixed the same way.
• And the space comes back. Deleted entries used to leave their space reserved forever — a Knowledge Base holding eleven entries could still occupy 150 MB, carried by every backup. LLypses now compacts the file after a large clear-out. Small deletions skip it, because rewriting the whole file to reclaim a few kilobytes would be slower than the problem.
• The folder tick box is visible. Every folder row has a tick box that selects or deselects everything inside it, which is what you want when you are clearing most of a Knowledge Base but keeping one folder. It only appeared when your mouse passed over it. Individual entry tick boxes still appear on hover.
• Built-in knowledge is part of the app, not part of your list. LLypses ships reference material about AI models and providers that keeps your members from giving last year's answers. It used to sit in your Knowledge Base as ordinary entries that could be edited, moved or deleted by accident — and if they were, your board quietly got worse with nothing to connect cause and effect. It is part of the application now, and updates when LLypses does. Anything you edited stays yours: LLypses can tell, and leaves it in your list.
• What it costs is said out loud. That material is sent to every member on every message, so it is part of what you pay on your own key — usually the largest single part. Roll Call now reports roughly what it costs, included in the total rather than quietly left out. There is no button beside it, because there is nothing for you to switch off.

Nothing here changes which models run or what they cost per message. Your members, your keys and your own entries are untouched.

The release where the app stops telling you things that are not true. Most of this one is about the gap between what LLypses showed you and what it was actually doing.

• The Knowledge Base can scroll again. With a large imported folder open the list simply stopped, and the entries past the fold could not be reached — so anything you wanted to delete down there was stuck. Fixed, and the same mistake is now checked for in fifteen other panels.
• Knowledge Base folders you can manage. Move entries into a new or existing folder, rename a folder, delete one. Deleting asks which you meant — remove the folder and keep the entries, or delete the entries too. Folders stay open or closed the way you left them.
• Select all looks like Select all. It was a small grey pill reading "All" that vanished the moment you selected anything, so it could never show you what it did.
• Two entries you were never meant to see. Every Knowledge Base contained two internal bookkeeping rows in a folder called _system. They are hidden now. Nothing was deleted.
• Adding a folder says what it is about to do. It used to walk everything, including dependency and build folders holding thousands of files that are not knowledge. It now skips those, counts what it found, and asks before committing — because a folder import cannot be undone in one step.
• The Roster tells the truth about your models. It warned that models would not respond when it had simply never tested them; it kept showing the previous answer after you changed a member's model; and it claimed a member would not reply while Auto was already answering for it.
• Check one key instead of all of them. Testing your keys checked every provider and every member. There is now a one-provider-at-a-time row, each showing its own request count.
• LLypses says when it has lost its own engine. The window used to look perfectly normal while everything quietly failed. One clear banner now, with Try again and Restart — and saving a custom model no longer blames your input for it.
• Scrollbars you can see. Some panels drew a bar nobody could see, some drew the thick Windows default. One consistent scrollbar now, at a contrast that meets the standard for a control.
• Room to work. Add Member collapses when you are not adding anyone, and remembers. Read the manual no longer looks identical to the button beside it.
• Sending work to the Workshop. The Directive field starts empty rather than pre-filling with the last thing said, and you can set more than one projects folder.

Nothing about your costs changed. No model was swapped, no default moved, and every check that spends a request still asks first and says what it will cost.

You can now choose who answers by clicking them, and one change affects what a message costs. Plus a cleanup for meetings that had quietly become expensive, and several things the app knew but never said out loud.

• Members in the sidebar are now buttons. Click to light one up, click again to turn it off. Whoever is lit gets your next message.
• This changes what an untagged message does. A message with no @tag used to go to @COO alone. It now goes to everyone currently lit, and each one is a separate call billed to your provider key. Your selection starts as whoever is already in the room you are working in, so nothing is lit that was not already there — but light eight members and send an untagged message, and eight members answer. Typing @COO still works exactly as before, and still overrides whatever is lit.
• Meetings were carrying hidden copies of themselves. If you asked for a session brief before version 0.6.5, that request was saved back into the meeting with a complete copy of the conversation inside it. 0.6.5 stopped new briefs doing this but did not clean up the ones already saved, so a room you briefed twice has been paying to re-send them with every message since. This release removes them on first launch and tells you what it recovered. On the worst example we have seen, nearly half of a meeting was these copies.
• Nothing is deleted permanently. A complete backup of your meetings is saved before anything is removed, and the app shows you where it is.
• A member that cannot see your picture now says so. Not every provider accepts an image. A member that received none, answering after colleagues who could see it, would assemble their descriptions into a confident first-person account of a picture it had never been shown. Members that cannot receive an image are now told so, and told to attribute what colleagues report rather than claiming the view as their own.
• Attachments are identified correctly. Every image was being labelled a PNG whatever it actually was, which worked because a pasted screenshot usually is one. Photographs are not. If nobody in the room can read the format, the board says so and answers without it instead of failing silently.
• The board no longer invents work it has not done. Fixed for a member's own actions in the last release; it now covers claims about other people doing work — the case that produced a 56-message account of an engineering team that does not exist.
• The Roster no longer clips on a smaller laptop. The Auto column was cut off with no scrollbar to reveal it.
• A licence key in the wrong box is reported for as long as it is wrong. It warned you once and then went quiet, leaving members failing with nothing pointing at the cause.
• A member you renamed keeps its name everywhere. If your own member shared a tag with a built-in one, yours answered but the built-in's name and colour appeared on the reply.
• Auto no longer picks a preview model for your whole board.
• Reporting a problem sends more of what we need. If the app misbehaves and you restart it before telling us, the record of what went wrong used to be erased by the restart. It is kept now. Nothing new about your conversations is included.

Two members of your board had been working without a row on screen. Plus a fix for meetings that were growing far larger than the conversation in them, and a full user manual.

• @CODEX and @INTEL were missing from the interface. Both were on your board the whole time — they answered when addressed, were counted in @ALL, and were billed to your provider key. They simply had no row anywhere you could look. @CODEX in particular was doing real work on every @ALL with no way to find out what it was, what it cost, or how to change its model. Both now appear in the Roster and the sidebar.
• Neither is added to your Inner Board by this update. Your @ALL costs exactly what it did before. Tick them in from the Roster tab if you want them.
• Session briefs were inflating the meeting they summarised. Asking for a brief saved the request back into the session it was summarising, so the next brief summarised a transcript containing the previous brief's prompt. One meeting reached 60,000 characters with its own transcript embedded twice — and because a session carries its whole history with every message, that was something you paid to re-send on every turn afterwards.
• LLypses has a user manual. Ten chapters and three appendices, organised by what you are trying to do rather than by which tab a button lives on. Linked from the Help tab.
• Help tab: a Core Members card listing who is on your board and what each is for, and the six quickstart steps now scroll to the section they describe.

The biggest release since launch, and most of it is repair. Thirteen defects were found by a deliberate hunt for things that looked right and were wrong.

• Exports were saving the wrong version. Exporting a meeting while it was still running saved the copy from when you opened it in Minutes, not the one on screen. Nothing warned you, because as far as LLypses was concerned it had succeeded. Word, PDF and Markdown all now read the meeting fresh.
• PDF export could fail on a tall diagram. A portrait flowchart could be too tall for the page, and the export failed rather than resizing it. Diagrams now scale to fit in both directions.
• A member could run a model you did not choose. If LLypses had never tested your chosen model against your key, it quietly substituted another — while the Roster went on showing your choice. It now distinguishes "we tested this and it failed" from "we have never tested this", and leaves your choice alone.
• Spend you were not being shown. A model with no known price recorded its calls as $0.00, permanently. The Usage tab now shows "unpriced" and says when the total is an undercount.
• Removing an API key now removes it. Clearing the field reported success, changed nothing, and left the key in use.
• Custom model endpoints work — they never had, in any released version.
• Long answers say when they were cut short, and an answer that never arrived no longer claims to have been.
• Auto is two controls, not one. MODEL is your choice; AUTO is Off, Optimized or Performance. Choosing one used to erase the other. Off is the default and nothing on your board moves when you install this.
• Minutes: tidy several meetings at once, grouped by week. Spend history is deliberately kept when a transcript is deleted.
• Tables now export as real tables instead of rows of |. Long code blocks get their own box with a Copy button. Ctrl+P prints the conversation instead of leftover diagram errors. The conversation scrollbar was always there and too dark to see.

Mostly customer feedback. A drawing you can actually use, a knowledge base that tells you it is working, and a board that stops agreeing the moment you push back.

• Diagrams you can use. Click any diagram for full screen — zoom, drag, Fit, then Copy, Save or Print. Saving gives the same quality whatever zoom you were at.
• Your knowledge base says when it is being used. Members had been instructed not to mention it, which made a working knowledge base indistinguishable from one that had not loaded. They now name the entry.
• The board stops caving. Members change their answer when given a reason — a source, an argument, a correction. Being contradicted is not a reason.
• Knowledge base entries export as Markdown, Word or PDF, and come back in.
• Offline says offline, instead of an error code.
• Misspelling a member's name no longer silently routes to someone else.
• Rooms can stay quiet unless spoken to.
• Model selection can follow your key — the first version of Auto.
• Report a problem from inside LLypses, in Updates & Feedback.

Your board can draw — and four things that were quietly wrong in every version before this one are fixed.

• Members can draw. Ask for a flowchart, a sequence, a state machine or a timeline and you get a diagram in the conversation.
• Exported documents were wrong in three ways. Found by opening the files rather than trusting that they saved.
• Nothing in a message reaches the internet. A message could reference a picture hosted elsewhere, and displaying it announced you to that host.
• Google changed its API key format to AQ. — the old AIza keys still work.
• Saving an API key could succeed without showing the key in the field.
• Only the member you asked for answers. Addressing someone outside your current room used to quietly involve others.
• The boardroom can hold a conversation. Four separate limits had prevented members genuinely responding to each other.
• LLypses now closes completely. Quitting used to leave part of it running.

Three things found within an hour of 0.6.0.

• The knowledge base could end up with two copies of itself if an older version started after an update.
• The What's new text showed scrambled characters.
• The tier dropdown in Roll Call was near-white on near-white.

• Roll Call — find out who can actually answer, before you ask.
• Refreshed knowledge base. The built-in AI landscape entries were written in May and were brought current.
• Retired models removed from the picker.
• Usage had been pricing four members against a model they never ran on.

• The Braintrust rooms were silent. Joining a Braintrust room and typing @ALL did nothing at all.
• Executives and Friends are now just Members. You no longer have to decide which kind of persona you are creating before you have written a word.

• Export saves the file. MD, DOCX and PDF in Minutes did nothing.
• @ALL means your room, not everyone you have ever created.
• Rooms have their own tab, instead of living three cards deep in Settings.
• Send to Workshop works on your machine — it had shipped pointing at a folder path that existed on exactly one computer.

Three things broken in ways you could not have worked around, and one request.

• Continue a past session from Minutes.
• Export to Word and PDF now works — it had been returning 404.
• @ALL respects the room you are in.
• Sessions no longer fail on unusual characters, such as a table drawn with line-drawing characters.

LLypses became a licensed product. This release turned on activation, and added a License card showing which of your PC slots you are using.

Reliability and cost.

• The boardroom keeps working when a model doesn't. One member's provider going down no longer ends your turn.
• Prompt caching — a large knowledge base is no longer re-billed in full on every message.
• Knowledge base size limit now applies to the total, not just per entry.
• Stop button fixed in two ways.
• The model picker tells the truth about what your key can reach.

• Two system tray icons became one. LLypses created a working tray icon in code while the configuration silently created a second, inert one.
• The greeting no longer appears twice on launch.
• Re-activating a PC you already activated no longer consumes a second slot of your three.

Fixes the boardroom going silent on free-tier Google keys.

• Members now use a Gemini model your key can actually reach — measured once when you connect a key, rather than assumed.
• The Gemini 2.5 models became paid-tier only in July 2026, and free keys got a 404 from every member. The shipped defaults no longer point at them.
• Provider errors are in plain English and name the fix.
• New Your Google Plan card showing your tier and what your key reaches.

The first release signed by Red Warden Studios LLC. Windows now names the publisher instead of reporting an unknown one. Also cleared retired image-model overrides that had been left saved against members.

Image generation was removed; LLypses became a text-only boardroom.

• Splash animation no longer sticks on L…
• Cohere errors on empty messages fixed.
• The knowledge base had been truncating long documents at 3,000 characters — raised to 50,000.
• A custom member whose tag matched a built-in no longer answers twice.

Image generation fixes, member defaults moved to gpt-4.1, expanded OpenAI model catalogue.

The rebrand. C-SuiteGPT became LLypses — correct spelling throughout, a two-phase animated splash screen, and new brand icons.

Uninstaller data choices moved into the app itself; no more second window.

Uninstall button added to Settings with a confirmation dialog, and a wider image-model selector.

The uninstaller's data dialog now appears before uninstall begins, and Cancel aborts cleanly with the app intact.

Improved uninstaller — choose which data to remove (API keys, sessions, knowledge base, settings, images) individually or all at once.

• Free Gemini onboarding for new users.
• Drag and drop files into the conversation and the knowledge base.

UX polish: priority button state, Target Project dropdown styling, a "new to API keys" guide in Settings, and a Getting API Keys card in Help.

Braintrust routing fixed, and a missing error handler that had been causing silent disconnections when an OpenAI request failed.

The Lounge Update. The boardroom got a backroom — a casual mode for thinking out loud with a smaller crew. Also added template rooms, custom model endpoints for self-hosted models, and upgraded every OpenAI-backed member.

Theme swatch and template card fixes.

DOCX and PDF export, scroll lock fix, knowledge base checkbox polish, @BRIEF and @VOTE, and DeepSeek, Perplexity and OpenRouter support.

Custom titlebar fixed — drag, minimise, close and maximise.

The first public release. Auto-updater pointed at the public releases repository.