Improved
- VibeManager now runs on our production servers. Your account, projects and chats come with you; you may be asked to sign in once more after the update.
What's new
Every release of the VibeManager app, newest first, Some notes are written for engineers.
The Windows app can be a version behind the Mac app. The download page always has the newest version for each.
New
- An Enterprise project has a Verifier: once the last piece of work in a part of the project is added to your code, it checks that part against your requirements, one scenario each, and the project's progress file shows which requirements are verified so far.
- With your own key, each role (planning, building, reviewing, verifying) can run on a model of its own.
- A budget per project, with limits on rework, rebuilds and review rounds, and on spend and repeated tool calls within one turn. When the budget runs out, the work under way stops.
- The project's checks run before every piece of work is added to your code, and on your main code after it.
- Planning documents get a fresh check before you sign them off, and the app asks for the sign-off itself.
- Agents look up current docs for the libraries they use, keep a list of the framework versions the project is on, and are warned when they use an API those versions no longer have.
Improved
- Team mode is now called Enterprise mode. Your projects carry on as they were.
- Every role is given the constraints you set, not only the Orchestrator.
- A reviewer's approval counts only when it cites the checks for that piece of work.
- Agents can no longer start a server that is open to your whole network.
- On Windows, the app installs in about half the time, from an installer a quarter smaller: it ships half as many files, without a second copy of the agents' libraries or 54 extra languages for the app's window.
- On Windows, an agent that has been idle for 10 minutes gives its memory back, and starts again where it was with your next message.
- On Windows, keeping the computer awake no longer runs a PowerShell window in the background.
Fixed
- A Quick chat no longer slows down and freezes while the agent thinks at length: the window rewrote everything the agent had ever written on each new word, and kept a copy of every step of the text it was writing.
- Dev servers and Docker stacks that an Enterprise build's agents leave running are stopped when the piece of work ends and when the app quits; on Windows too, where a dev server you started yourself is left alone.
- Every question the app asks you can be answered in your own words under Other…, sign-offs too.
- An answer reaches the teammate that asked it, and a member that stops after it reports no longer reads as a crash.
- When the build can't read a plan, it says what is wrong with it.
- Agents are kept from reading your secrets by any route, but no longer refused for text that merely names one.
- On Windows, stopping a Quick app's preview also stops its dev server, which kept port 3000 and its memory.
- On Windows, an Enterprise build's checks no longer wait forever when they leave a server running.
- On Windows, re-checking a security fix looks only at the folders it changed, not the whole project.
Improved
- In every chat, scrolling up to read earlier messages takes the message box down to one row and shows Jump to latest, as Quick mode already did. Sending a message brings the chat back to the end.
Fixed
- Scrolling up a chat with a trackpad no longer shakes it or snaps back to the end, also while the agent is still replying.
New
- Settings > Appearance > Reduce transparency makes every panel solid instead of see-through. It is on by itself when your Mac asks for less transparency, and always on Windows.
- In a security audit, click the Auditor or any agent to follow it in the side panel: the Auditor's plan, who is working now and what it decided; an agent's brief, its steps and what it ran in the sandbox. Open in full shows it as a page.
- Select findings to get a bar with Export and Fix with the team. A finding the agents could not confirm asks for your look: keep it, dismiss it, or open the code.
Improved
- A new look across the app: panels like thick glass with a fine edge, one blue for buttons and links, a white page and a soft grey sidebar in light, charcoal in dark. The theme picker is gone; Light, Dark and System stay.
- The side panel's tabs are tidier, and Quick mode's progress shows as a small moving mark.
- Security has a new home: your audited projects in one list, each opening a page with its last audit and its history. Choosing an audit's depth shows the team each depth sends.
- A running audit shows each agent at work and folds the finished ones away. Each agent has a small face that moves only while it works.
- Findings show their severity as signal bars in distinct colours, and the details say how each one was checked, with the suggested fix.
New
- VibeManager runs on Windows 10 and 11, on Intel/AMD and Arm PCs. The installer is not signed yet, so Windows shows a SmartScreen warning the first time: choose More info, then Run anyway. The Windows app does not update itself yet; install the next version over it.
- On Windows, agents run their commands in Git Bash. When Git for Windows is missing, VibeManager offers to install it and carries on once it is there, without a restart.
- On Windows, the Terminal opens PowerShell in the project folder, Open in lists Explorer and the editors you have, and the window controls follow the theme. A security audit on Windows reads the code only: running it in a sandbox needs macOS.
Fixed
- In Team mode, Save the plan now saves a plan kept in a new folder; before, the build kept asking for the plan.
Earlier releases
New
- Quick mode. A new project asks Quick or Team: Quick builds a whole app from what you describe and shows it running in the Preview as it goes. A Quick project stays Quick; Keep building asks for the next change.
- Standard or Smart, in the composer, for each Quick project. Standard is quick and low-cost and right for most apps; Smart takes more care on harder apps and costs more.
- Before a new app starts, a short brief card shows what will be built: what the app is, its pages, how much it needs (a simple site, a server, saved data or sign-in) and a choice of looks. Change anything, then build.
- While an app is built you see it live: a status line from the moment you send, a progress bar of your requirements above the composer, and the work grouped under the requirement it serves. Each request ends with a card that says what was built and checked, with a picture of the app.
- When the app needs your decision or a key (a payment provider's, say), it asks in the composer, and the key goes into the project without passing through the chat.
- In the Preview, pick an element of your app and say what to change about it. View as shows the app as a visitor, a user or an admin.
- Apps can store uploaded files, run scheduled jobs and use AI. Signed in, an app's AI runs on your VibeManager account; no key needed.
- Apps get guides for payments, maps and notifications, five design presets, and a design check that flags generic layouts.
Improved
- The Graphite theme is redone, and the sidebar is narrower.
Fixed
- A chat no longer jumps to its top when a reply arrives.
- Quitting VibeManager waits for open terminals to close, so it no longer crashes on the way out.
New
- Settings > Privacy. "Help improve VibeManager" lets you choose whether we may use your prompts, agent conversations and code excerpts to improve VibeManager; it starts off in the EU and EEA, the UK and Switzerland. The page links to our privacy policy.
- The sign-in screen says, in one line, what VibeManager sends to our servers and how long we keep it.
Improved
- The AI model behind each agent and each audit step is now chosen by VibeManager's servers, so we can move a step to a better model without an app update. The app no longer lists model names or prices, and the cost it shows is what the call was billed.
- An audit keeps the models it started with, even after you pause it, quit the app or resume it later.
- When we move a step to a different model, an agent already running switches at your next message, never in the middle of a reply.
- To find and fix problems, the app sends VibeManager a record of each session, agent step and tool call. We keep these records for 30 days; Settings > Privacy and the privacy policy say what's in them.
Fixed
- An agent could be stopped while a helper task it had started in the background was still working.
New
- Back and Forward arrows beside the sidebar toggle step through the pages you have visited, as a browser's do. ⌘[ and ⌘] do the same; change them in Settings > Keybindings.
Improved
- Security audits follow a fixed plan of steps (map the code, look for weaknesses, check each lead, double-check, write the report) instead of one agent managing a team. They finish much sooner and cost far less: a standard audit of a medium project now takes under an hour.
- An audit picks up where it was after you quit the app, after a crash, or when you resume a paused or stopped audit. Work already finished is not done again.
- Stopping an audit writes its report straight away, marked incomplete, with the leads not yet checked listed.
- The VibeManager mark is no longer in the title bar; the arrows take its place.
Fixed
- An audit could end by itself partway through, with its checks still running. It now runs to the end.
- An audit that fails stops using the model at once.
Improved
- An update downloads when you click it, not by itself: the sidebar's update button, Settings > About, the sign-in screen and Check for Updates… all offer Download, then Restart to update.
- If macOS asks whether VibeManager may use "VibeManager Safe Storage" the first time you open this version, choose Always Allow. It keeps your sign-in encrypted on this Mac, and it should not ask again after updates.
New
- VibeManager > Check for Updates… in the menu bar, which says whether you are up to date or an update is on its way.
- The sign-in screen shows an update that is downloading or ready, with a button to restart into it, so an old version that can no longer sign in can still update itself.
- A new Security intro video, 1:04, with its sound.
Improved
- A new look for the installer, with the team waiting for your first app.
Fixed
- Sign-in works again: the app was looking for our server at an address that no longer answered.
New
- VibeManager updates itself: it checks for a new version, downloads it in the background and tells you when it is ready, with what's new; you choose when to restart. This needs the app in your Applications folder and a release signed by us. Builds before the first signed release can't update themselves, so install that one from its DMG once.
- The app is signed and notarized by Apple, so it opens without the first-open warning.
The first numbered release. The 0.1.0 DMGs shared before it were test builds; their projects carry over.
New
- A chat app for your projects: projects and their sessions in the sidebar, one conversation with the Orchestrator per session, the specialists it brings in shown as rows in its replies, and each specialist's conversation, a list of changes or the preview in a panel on the right. Board, Preview and Documents are tabs in the project header.
- A New project dialog with a name, a folder and an option to track changes with git, and a first-run welcome.
- Sign-in with your VibeManager account in the browser (invite only for now). The app carries no model key: every agent calls the model through our server.
- Settings for your key, the appearance, GitHub and keyboard shortcuts, and a command palette on ⌘K.
- Changes you can see and commit: a panel for uncommitted changes, a piece of work's changes, each attempt and what was added to your code; the branch and its changes under the chat box; a Commit dialog that never offers a `.env` file.
- Each piece of work is built in its own copy of the project, inside the project's folder, and added to your branch only once it is reviewed and approved. Nothing is written outside the project before then.
- The team follows a proven planning and build method that comes with the app: the Orchestrator picks the steps a request needs, the specialists write the brief, requirements, UX design, architecture and pieces of work, and a Developer and a Reviewer build and review each piece of work, up to three rounds.
- The Orchestrator asks you to sign off the requirements, the architecture and the breakdown into pieces of work before it moves on, and asks again if a document changes after you approved it.
- Questions a specialist cannot answer reach you in the chat box as a decision, with the recommended option marked.
- Several sessions per project, sharing the project's specialists.
- DMGs for Apple silicon and Intel, ad-hoc signed, with everything the team needs inside the app.
- Security: choose a folder and a team of agents checks it for security problems. They read the code, try each weakness in an offline sandbox on your Mac and double-check it before it reaches you; nothing in the folder is changed. Watch the team at work, see findings in a table or a board, export them as PDF, Markdown, SARIF, JSON or CSV, and hand them to your project's team with Fix with VibeManager. A Re-check after the fix marks them Fixed. A short intro video on the Security home shows how it works.
- A finding is kept as one issue across audits, so a re-check knows it even when it is worded differently. Accept Risk and Not a Problem ask for a reason, and the finding keeps its history.
- Settings → Storage: how long an audit's working files are kept (7 days by default), the space each audit uses, and Prune Now. With Keep finished reports in your VibeManager account on (off by default), finished reports are kept in your account; your files and the agents' transcripts never leave the Mac.
- A panel on the right with Files, Preview, Team and Terminal tabs; links, files and folders in a reply open where they belong.
- Side chats, Send later, a choice of what a follow-up does while the team works, your own instructions for every agent, a limit on how many agents work at once, notifications when the app is in the background, Keep awake while agents work, retries when the AI service is busy, and the window opening where you left it.
Improved
- Every agent runs on the same AI model; there is no model picker.
- Pieces of work that do not depend on each other are built two at a time, and a piece gets a thorough review only when the plan marks it high risk, so a plan is built faster.
- The spend shown in the app is what each call was billed, including the helper agents a step starts, and a long or failed turn is counted too.
- Specialists' words, build rows and the plan card are shown in plain product words; an error says what went wrong in one line, with the details behind a disclosure.
- A build that waits for you shows its card once, pinned above the chat box, so its buttons stay still while the conversation streams; the conversation keeps a one-line note of it.
- The DMG window and its "Read me first" say how to open the app the first time on macOS 15 and later, and on macOS 13 and 14.
Fixed
- Deleting an audit removes everything it made: its folder, its conversation and every trace in the app's database.
- The agents of an audit read only the folder being audited and their own files, nothing else on your Mac.
- Each signed-in account has its own projects and audits; another account on the same Mac starts empty.
- Starting the same audit twice, or a second Fix while the first is still running, no longer mixes them up.
- A sign-off is kept when a specialist only updates a document's status after you approved it.
- A quick review no longer fails working code over minor findings.
- A message sent before a key is set says so, instead of reading as an internal error.
- The note that a build was added to your code shows in the session.
- Restoring the demo project never races another run, and keeps the project's history and documents.
- A piece of work whose changes conflict with another piece added just before it is built again on your updated main code, even after its last review round, and that round does not count against its three attempts. Try again on a conflict builds the piece again on your branch as it is now, not the same changes again. A conflict with your own commits still waits for you.
- A long decision title stays inside the decision's panel.