What changed.

Everything below is in 0.0.54. Updating once brings all of it.

Fixed

• Long conversations that contained large images no longer get stuck. A big picture in the history could push a thread over its memory budget, making it rewrite its summary forever: replies were cut off after a single word and new messages got no answer. Cleanup now knows how to let go of oversized pictures, and reading very large image files is refused with advice to resize instead, so a thread can always recover on its own.

Everything below is in 0.0.53. Updating once brings all of it.

Fixed

• Conversations with images in them no longer break. Sending a message in a thread that contained pictures, including ones Cyborb generated itself, could fail with Service Unavailable on every try, cut replies off mid sentence, or answer one message late. Each picture is now understood once and remembered, instead of being re-processed on every message.
• Generated images are now used, not just made. When Cyborb creates a logo or other asset while building your project, it wires the file into the project before finishing, or removes it and tells you why.
• The message box no longer draws its buttons on top of the placeholder text when the window is narrow.

Everything below is in 0.0.52. Updating once brings all of it.

Fixed

• Messages no longer disappear after a reply. On some OpenRouter models, especially free ones, a conversation could wedge: the answer appeared, the thread kept saying Working, and your next message showed as sent but never got a response, or got one message late. Cyborb now closes a turn as soon as its reply is fully delivered, so every message you send starts fresh and gets answered. A stuck Working state also ends on its own within seconds instead of minutes.

Everything below is in 0.0.51. Updating once brings all of it.

New

• Change the model mid conversation on OpenRouter. You no longer need a new thread to try a different model. Pick another model in the composer at any point and the same conversation continues with it, full history included. If a model refuses or is unavailable, switch to another and keep going.

Everything below is in 0.0.50. Updating once brings all of it.

Fixed

• Signing in works again. A recent build changed where the browser returns after you approve sign-in, and our sign-in service had not been told about the new address, so it refused with a red error instead of signing you in. The address is registered now, existing installs are already unblocked, and a release check makes sure this cannot ship broken again.
• Kimi tells you when it has nothing to say. A conversation could end silently with a blank thread. Cyborb now notices an empty reply and says so, instead of leaving you staring at nothing.
• Messages name the model you chose. Threads running through OpenRouter showed system text that said Claude even when you had picked a different model.
• Rate limiting is explained in plain words. When a provider slows you down, Cyborb now says that is what happened instead of showing a raw retry message.
• OpenRouter works with every model. Models outside the Anthropic family could fail with a request error caused by a tool feature they do not support. Cyborb now turns that feature off for them automatically.
• The thinking indicator is visible on the light theme. It was a pale dot on a pale background; it now darkens to match the theme.

New

• Install a missing provider in one click. If a provider's command line tool is not on your machine, its card now shows an Install button. Click it to see the exact command, and run it from right there. Cyborb never runs anything you have not seen.

Everything below is in 0.0.49. Updating once brings all of it.

Fixed

• Every OpenRouter model is now listed. Earlier builds showed only a handful. Cyborb now reads the model list from OpenRouter directly, so new models appear on their own, and you can still add any model by name.
• OpenRouter and Kimi now show their real logos in their own colors.
• OpenRouter no longer depends on how Cyborb was started. It now records the exact location of the ori tool when you set it up, instead of looking it up each time, so it keeps working whether you open Cyborb from the Dock or a terminal.
• OpenRouter conversations no longer break part way through. Sending a message could fail with an internal mismatch error.

Setup

OpenRouter runs through OpenRouter's own ori tool. Install it once, run ori login, and Cyborb picks it up:

``
curl -fsSL https://openrouter.ai/labs/ori/install.sh | bash
``

If you would rather not sign in with ori, you can paste an OpenRouter API key in the provider's settings instead.

Everything below is in 0.0.47. Updating once brings all of it.

Fixed

• OpenRouter works. In 0.0.46 it appeared in the provider list but failed on every message with a mismatch error. That is fixed, and it is worth updating for if you tried OpenRouter and it did not work.

Changed

• OpenRouter now runs through OpenRouter's own ori tool rather than being wired up by hand. Install it once with the line below, run ori login, and Cyborb picks it up. Nothing to paste, and your sign-in stays with OpenRouter rather than being copied around.

``
curl -fsSL https://openrouter.ai/labs/ori/install.sh | bash
``

• If you would rather not sign in with ori, you can still paste an OpenRouter API key in the provider's settings.
• OpenRouter and Kimi now use their real logos.

Everything below is in 0.0.46. Updating once brings all of it.

New

• OpenRouter. Add your OpenRouter key in Settings under Providers and you can work against hundreds of models on one balance, including Claude, GPT, Gemini, DeepSeek and Kimi. Any OpenRouter model name can be added by hand, so you are never waiting on us to list a new one. If you already use OpenRouter's own ori tool, Cyborb picks up the key it left in your environment and you paste nothing.
• Kimi. Moonshot's Kimi Code CLI is now a provider in its own right. Install it, run kimi login once, and Kimi appears with its models ready. Kimi models are also reachable through OpenRouter if you would rather use one balance for everything.

Notes

• Both providers are marked Early Access while they settle.
• Kimi's own service was returning errors from Moonshot's side while this was built. If Kimi replies with nothing, that is their service rather than Cyborb, and the same models work through OpenRouter in the meantime.

Everything below is in 0.0.45. Updating once brings all of it.

Fixed

• Providers no longer say they failed when they are fine. Claude, Grok and Cursor could report "installed but failed to run" right after Cyborb started, and then stay that way for five minutes, even though the tool on your machine was working. Cyborb was judging each one on a single check taken while everything was still starting up. It now checks again before deciding, so a slow moment during startup no longer looks like a broken install. A tool that really is broken is still reported, with the same message as before.
• Library, Schedules, Usage and My sites open beside your chat. Choosing one of them from the sidebar used to take you out to Settings and lose your place. They now open in the panel next to your conversation, so you can look something up without leaving what you were doing.

Under the hood

• Site previews are only loaded over a secure connection, so a preview can never be pointed at a file on your own machine.
• Updates are only accepted from a secure connection.

Everything below is in 0.0.43. Updating once brings all of it.

New

• An asset library. Every image, video, track and 3D model you have generated is now in one place, searchable by the prompt that made it, in Settings under Library. It fills in from your whole history, so anything you made before this update is already there.
• Pin chats to the top of the sidebar. Right-click any chat and choose Pin to top. Drag pinned chats to put them in the order you want.
• Choose what lives in your sidebar. Library, Schedules, Usage and My sites can each be shown as a sidebar entry instead of living only in Settings, and you can drag them into the order you prefer. Nothing changes unless you turn one on.
• A warning before you run out. Cyborb now tells you when you have used 80 percent of your budget, in the chat where you are working, instead of staying quiet until you hit the wall.
• Video and 3D generation asks first. Generating a video or a 3D model can use a large share of a weekly budget in one go, so Cyborb now asks before it does either. Images are unchanged and still generate without interruption.
• Site analytics in the app. Visitors and traffic for your published sites, in My sites.

Fixed

• Restarting for an update no longer throws away work in progress. If the agent is still working when you restart to install an update, Cyborb tells you what is running and lets you wait. The update stays downloaded either way.
• The first-run screen fits your window. The agreement and the buttons are always visible, so you no longer have to scroll to find how to continue.
• Claude Opus 5 and Claude Sonnet 5 are selectable.

Everything below is in 0.0.41. Updating once brings all of it.

Fixed

• Signing in with Google works again for new accounts. Choosing Continue with Google as a first-time user ended with "OAuth callback failed" and no account was created. Existing accounts and email sign-in were unaffected, which is why it went unnoticed for a while. The cause was a bug in the sign-in library Cyborb bundles, which treated the ordinary first-time step of "this Google account has no Cyborb account yet" as a failure instead of the moment to create one. That library is updated.
• Sign-in errors now say something useful. If a Google sign-in does not finish, Cyborb tells you what happened and points you at email sign-in, instead of showing an internal error code.

Everything below is in 0.0.40. Updating once brings all of it.

New

• Claude usage in Settings. If you use Claude on a subscription, Settings now shows how much of your 5 hour and weekly limits you have used and when each one resets, next to your Cyborb plan and your Codex usage. Nothing is shown for accounts on an API key, AWS Bedrock or Google Vertex, because those have no plan limits to report.
• Claude Opus 5 and Claude Sonnet 5 are selectable. The picker previously stopped at Opus 4.8 even though newer models had shipped.

Fixed

• Linux builds could be assembled from stale files. Our build service caches the agent files it downloads for four hours, so a Linux build could quietly be made from the previous copy, and a change to the built in skills might not reach it. Every Linux build now fetches fresh. The 0.0.40 AppImage is confirmed clean.

Everything below is in 0.0.39. Updating once brings all of it.

Fixed

• Claude is selectable again. In 0.0.38 Claude signed in correctly and then sat greyed out in the model picker, labelled Limited, so you could see it was authenticated but could not use it. That was our bug: the subscription warning was being carried in the field that decides whether a provider can be picked at all. Claude is usable again, and the warning still appears on its card in Settings where it belongs.
• Three more long dashes removed from text you actually see, including the model picker tooltip and the risk note at the top of the Providers panel.

Everything below is in 0.0.38. Updating once brings all of it.

New

• Claude works with the account your Claude CLI is already signed in to, including a Claude subscription. Please read the next paragraph before you rely on it. Anthropic does not permit third-party apps to use Free, Pro or Max plan credentials, and they can restrict your Anthropic account for it without warning. Cyborb no longer blocks it, but the choice is yours and so is the risk. Cyborb shows a warning on the Claude provider whenever it is running this way. To avoid the risk entirely, add an Anthropic API key in Settings, Providers, Claude, or configure AWS Bedrock or Google Vertex.

Fixed

• The first screen follows your theme. The Terms screen shown on a fresh install was always dark, so a new install opened black and then turned white once you were through it. It now matches whichever theme you are on, which for a new install is Light.
• Cleaner writing throughout. Ten places in the app used a long dash mid-sentence. They now read as ordinary sentences.

Everything below is in 0.0.37. Updating once brings all of it.

New

• Cyborb opens in Light. A fresh install now starts on the light theme instead of a dark one, on every machine, whether or not your system is set to Dark Mode. If you have already picked a theme, you keep it. If you never picked one, you can change it any time in Settings, Appearance.
• Usage for your other providers. If you use Codex, Settings now shows how much of your window you have used and when it resets, next to your Cyborb plan. This appears only where the provider actually publishes that information: Codex does, and Claude, OpenCode and Grok do not, so nothing is shown for them rather than an empty meter.

Fixed

• Cyborb Glass is one surface again. The bar along the bottom of the chat was painted a different tone from the chat above it, which put a visible seam across the window in the one theme whose whole point is that it is see-through. That band is gone. The chat, the area behind the composer and the strip below it are now the same glass, and the box you type into is the only thing tinted, so what you type stays readable.
• Source control no longer looks broken. If you use GitHub and nothing else, the panel used to lead with three failures for GitLab, Azure DevOps and Bitbucket, none of which you had asked for. It now lists only the forges you actually have, and folds the rest into a single line you can open if you ever want to add one, with what each of them needs.

Everything below is in 0.0.36. Updating once brings all of it.

New

• My sites shows a picture of each site. Every published site now has a real capture of its page on its card, so you can find the one you want by looking instead of by reading URLs. The picture is taken on your machine, cached, and refreshed when you republish.
• The composer tells you which project you are in. The bottom left used to say "Local checkout" and nothing else. It now shows your project's name, and clicking it lists every project you have plus a way to add another.
• Claude can be set up inside Cyborb. Settings, Providers, Claude now has a field for an Anthropic API key. Create one at console.anthropic.com. An AWS Bedrock or Google Vertex setup works too.

Fixed

• Cyborb no longer talks unless you ask it to. A second speech setting sat one row underneath the Anima voice switches, and turning it on made the ordinary chat read every reply out loud, with nothing on screen to say why and nothing in the chat to stop it. Speaking is part of Anima Mode now and nowhere else. If you had that setting on, it is cleared for you on update. Anima itself is unchanged.
• Claude showed as Unavailable and suggested a fix that could never work. The message was generic text shown for every provider, telling you to sign in through the terminal, which is the one thing that cannot work for Claude here. Each provider now gives its own guidance. Note that a Claude Pro or Max subscription cannot be used in Cyborb: Anthropic does not permit third-party apps to sign in with subscription plans, and doing it can get your Anthropic account restricted.
• Glass mode looks like glass again. The bar at the bottom of the chat was a solid dark navy slab under a grey transcript. The dock is transparent now, and only the box you type into is tinted, enough to keep your text readable over whatever is behind the window.
• The voice waveform in Anima Mode kept animating while Cyborb was behind another window. It stops when you cannot see it.
• Site preview captures are no longer counted as visitors in your own site analytics.

Everything below is in 0.0.35. Updating once brings all of it.

New

• Cyborb Air is now the model a fresh install starts on. Before this, the app opened on whichever other CLI you happened to have installed.
• The sidebar has a New chat button, and opening the app drops you straight into a chat instead of a "pick a thread" screen.
• My sites is rebuilt. Every site is a card now, showing its live URL, size, file count and when you published it, with traffic one click away.
• Your plan in Settings is a card you can tilt and hold.
• Usage now tells you how long until each window resets, instead of only how much you have spent.
• Claude can be used with an Anthropic API key entered directly in its provider settings. Anthropic does not permit apps like Cyborb to use a Claude Pro or Max subscription sign-in, so a key from console.anthropic.com is the supported route. AWS Bedrock and Google Vertex setups work too.

Fixed

• Updating no longer loops. If a newer version appeared while an update sat waiting for you to restart, the app installed the old one and immediately asked you to update again.
• Anima stopped asking everyone to pick a template. She asks a question when she actually has one.
• In Cyborb Glass, the chat area was a flat grey panel. It is transparent again, the way the theme intends.
• Grok and OpenCode report their real state. A working Grok used to say authentication could not be verified, and a signed-out one told you to check server logs; both now say what is true and what to run.
• Settings no longer lists things that do not exist yet.

Faster

• The app starts noticeably quicker. It was reading your shell profile twice before showing a window, once in the app and again in the backend.
• Long replies with code stay smooth. Code blocks were being re-highlighted from scratch on every token, which got slower the longer the block grew.
• A minimised window now costs almost nothing. Several animations kept drawing at full speed while you could not see them.
• The git panel stops re-asking GitHub about pull requests every thirty seconds.

This release covers 0.0.31 through 0.0.34. If you are on anything older, updating once brings all of it.

Fixed

• The unrecoverable "Something went wrong (fetch-session-state)" launch error is gone. It could hit anyone who reloaded the app or signed in on a second surface, and no amount of retrying fixed it. The app now issues itself session credentials that survive the whole run.
• Signing in closes its own window when you finish instead of leaving you on a spinner.
• The false "You're signed out" banner now repairs itself: if you are actually signed in, the app reconnects your session instead of asking you to reload.
• The first-run terms screen no longer clips its top edge on smaller windows.
• The composer no longer reserves a block of empty space under your text.

Anima

• Listening actually listens. The voice detection threshold was tuned against real microphones; before this, the waveform moved but Anima heard nothing.
• When Anima hears silence, she says so instead of going quiet.
• Her voice no longer dies after the first reply of a session.
• Her internal instructions no longer appear as a message in your threads.
• The Speak and Listen controls sit on one line, properly aligned.
• The scripted setup questionnaire is gone. Anima asks a question when she has one, not because a wizard told her to.

Under the hood

• Session tokens refresh reliably while the window is in the background.
• Audio playback no longer silently breaks when the system output device changes mid-session.