RU Download
Esc
What’s new

Version
0.24.0

The app updates itself: an “Update” button appears when a new version ships. Here is what changed.

Download the app
0.24.0

One file and the computer is ready: a Windows installer, everything installs itself

  • Windows: a regular VibeBot-Setup.exe installer instead of an archive. It installs into the user folder without an administrator password, adds Start menu and desktop shortcuts, launches the app and uninstalls from “Apps & features”. The app finally has its own icon.
  • Everything needed installs itself, with no “Install the tools?” question: Node, Python 3.12, git and code search come from the installer itself, even offline. ffmpeg, pandoc and the bot’s browser are fetched in the background once the computer is connected to the swarm; turn it off in the tray menu (“Fetch the bot tools automatically”).
  • On a clean Mac the bot no longer hits Apple’s “install developer tools?” dialog: git is now our own portable build. Python comes from our server rather than GitHub — more reliable from Russia.
  • Ollama installs into the workshop with one click — no installer, no administrator. The app starts the model server itself when the swarm asks for models from your computer.
  • Mac: an app launched from the disk image or from Downloads offers to move itself to Applications — one click, no dragging. The image background says what to click if macOS refuses to open it.
  • Permanent build addresses: /download/latest/win-setup.exe and /download/latest/mac-arm64.dmg always point at the current version — for emails, docs and scripts.
0.23.0

The bot works on websites the way you do: with a mouse and a keyboard

  • The bot no longer merely “reads pages” — it works on them. Real mouse clicks and real typing: fields that validate input (VIN lookups, React forms) finally accept what it types, and icon buttons without a caption can be clicked.
  • The bot sees the page as a list: every field, button and link has a label, a name and a state. It used to guess where to click and miss — now it takes the label from the list. The answer also became four times shorter: fewer tokens per step.
  • Blind clicking is no longer possible: if a popup or a banner covers the target, the bot says “the element is covered” instead of reporting success. A false “done” is the most expensive mistake, because you believe it.
  • Tabs, system dialogs (“Delete the item?”), downloads into the working folder and uploading your files into site forms all work now. A link that opens in a new tab is no longer lost.
  • You sign in yourself: the bot opens the sign-in page and stops, you press “Take control” on the “Swarm computer” page and type the password. Passwords and SMS codes are never passed to the bot — it continues in your session.
  • Placing an order, paying, deleting and accepting terms are always confirmed — in any conversation mode, including the autonomous one. Signing terms is a deal, and it asks separately, naming the document and the site.
  • A new “Browser” tab on the “Swarm computer” page: you can see what the bot clicked, on which sites and how it ended. Values typed into fields never reach the log.
  • The bot no longer refuses work it can do. “I cannot interact with the site” while the tools are right there is now flagged the same way a false “done” has always been, and the bot is told to try.
0.22.0

The bot works alone or hands the work to the swarm — your call

  • A third button appeared next to the input: “Alone”, “As a swarm” or “By the task”. In swarm mode the bot splits the work into assignments, hands them to other bots and then accepts the results; in “By the task” it decides for itself — small work alone, large work handed out — and always explains the decision in the chat.
  • The swarm finally works in parallel instead of in turns: independent parts run at the same time, dependent ones wait. Two assignments run at once by default — exactly what the local model server can take — and they are grouped on one model so the server does not swap weights between turns.
  • A whole piece of work is now an order: the chat shows a single card — the goal, who does what, what it cost and the outcome. It updates in place instead of flooding the feed with “assignment sent”, “assignment returned”.
  • Workers return results as data, not retellings: a short summary and links to what was made — a file path, an address, an image name. What one bot finds goes on the shared order board, so the next one does not look for it again.
  • A new “Ops room” section: a board with four columns — new, in progress, up for acceptance, closed. From there you can accept work on the lead’s behalf, send it back with a note, reassign it, withdraw an assignment or stop the whole order.
  • An order has a cap — 400 thousand tokens and two hours by default. Once it is spent, the work stops honestly with a summary of what is done, instead of quietly burning a night of GPU time.
  • You do not have to name the worker: the swarm picks one itself — by role, by what the bot has already learned, by how busy it is and whether its model is available — and says why it chose them. If a bot goes silent, the work moves to another one once.
  • In the morning a single digest arrives: what was done overnight, what is in progress, what stalled and what needs you. Instead of a stream of notifications.
  • Handing work out grants no new rights: bans, tool permissions and approvals work as before, and the worker runs in the same project and under the same approval rules as the order itself.
  • Pressing “Update” no longer ends with an error dialog. The new version installed even before, but the outgoing copy of the app crashed on its already-closed window when the freshly started one asked to show it — so the update looked like a failure.
0.21.0

The swarm sees where it misses, and adjusts

  • The “Quality” page gained breakdowns: you can see which bot and which model reach for the wrong tool most often, and in what kind of work. Before there was a single number for the whole swarm.
  • Outside connections now explain themselves like built-in tools: in a connection’s card you can write when to use each of its tools — “in this task database, write only to the ‘Work 2026’ project”.
  • Paid services are marked: the bot knows that search on your key spends your quota, and that Stripe or PayPal calls are real money, so it reaches for them only when there is no free path.
  • Turns the owner did not start — a routine, an email, a handed-over task — are no longer interpreted from your last chat line from a week ago: the bot looks at the reason for the turn.
  • The swarm tunes which tools it shows: a set that has been opened for nothing for a month stops opening; a set the bot kept opening by hand opens right away. All of it is visible on the page and undone with one button.
  • Prompts became targeted: a bot that picks the right tool nine times out of ten no longer gets the obvious repeated to it, while one that misses gets more detail. You can pin the level yourself.
  • Before a release, a tool-choice acceptance run is executed: seven trap tasks where the workaround also produces a result. A lost task stops the release.
  • The Mac and Windows app now updates a new way: when only VibeBot’s own code changed, it downloads two megabytes instead of a hundred and seventy — seconds instead of minutes.
  • A dropped connection no longer restarts the download: it resumes where it broke off, and a silent server gives a clear error in fifteen seconds instead of hanging.
  • If a new version will not start, the previous one comes back by itself within half a minute — and says why. You can also go back by hand, with a button in “About”, where the update log is now visible too.
  • Choosing “Quit” no longer brings the app back: with an update ready, quitting used to reopen it half a minute later.
  • Every release is now signed with our key: the app will not install an update that was not published by us, even if the whole server were replaced.
  • In “About” you can see which version each of your computers runs, when it last checked for updates, and how the last attempt ended.
  • The chat no longer goes completely quiet: stay idle for a minute and a town of tiny bots comes to life behind the conversation. They raise buildings, carry crates, weld seams and now and then square off in a mock scuffle. The first mouse move fades it out. Turn it off with “Living background” on the “Appearance” tab.
0.20.0

Bots reach for the right tool

  • Every tool now says when to use it and when not to: “not for a one-off command — use shell for that”. The bot no longer picks a tool because its name sounds close.
  • Before working, the bot works out the kind of work — project, files, mail, images, an outside service, your computer — and opens the tools it needs right away, without spending a step on it.
  • Skills you saved now apply themselves. A skill used to work only when you typed “/”; now the bot finds the fitting one from your request and works the way you wrote down.
  • If the job needs a service you have not connected, the bot does not build a workaround through the terminal — it names the connection that fits and offers to add it. You are still the one who installs it.
  • What cannot work is no longer run just to fail: no computer of yours online, no Studio, no plan — the bot says so at once and suggests what to do instead.
  • The “Quality” page gained a second number — right tool first time: the share of turns where the bot never reached for the wrong thing. Next to it you can see which tools are taken for nothing most often.
  • Every action in the chat now has a “wrong tool” button: press it and the bot remembers your call for that kind of work. It becomes its lesson, not a wish.
  • The swarm exam can now catch “done with the wrong thing”: five new tasks where the workaround also produces a result, but the tool is what gets checked.
  • A new button under the input box, “How the bot works”: ask when it matters (as before), work on its own without asking, or get your approval for every action that changes something. It switches mid-task and takes effect from the bot’s next action.
  • The mode can be set for the whole swarm on the Swarm page, which also shows the conversations where you chose otherwise and lets you undo that there. Bans stay bans: commands that change the whole system will not run in any mode.
0.19.0

Projects: snippets, archives and honest package updates

  • A start-up error and a red check are shown right next to the line of the file: click and the caret is there. Previously “Price.jsx:31” had to be carried by eye from the terminal to the editor.
  • A snippet library appeared: a booking form or a site header made once can be dropped into any other project instead of being rewritten.
  • An existing project can be brought in as an archive (zip or tar.gz) — it unpacks itself — and a new one can start as a copy of an existing project: files are taken, the history stays its own.
  • “Update packages” runs the checks before and after and says plainly if the update broke the project. A snapshot is taken first.
  • You can see how much memory and CPU each project process eats, and we warn in advance when the swarm computer is close to its memory limit — otherwise the system kills the dev server silently.
  • A one-off half-hour link opens the project preview on your phone without the swarm password — to look at the site for real, with touch and a keyboard.
  • History gained search by the line of code itself (“when did this address get here”) and viewing files of any version without rolling back.
  • In the project terminal, Tab completes a path just like a normal shell.
  • A project now has its own log: who did what to it, without actions from neighbouring projects.
  • If a published site goes down, you hear about it by email and in the connected channels rather than from a client. Email matters because the server cannot reach Telegram — there is no outbound proxy in the container network.
  • Editing a very long file no longer lags: the visible window is highlighted instead of the whole text.
0.18.0

Projects: a workplace, not a showcase

  • The Studio has arrived: the bots can draw. Ask in your own words — the picture lands right in the conversation, not as a link to someone else’s service.
  • Your own ComfyUI does the work: on this server, on another server by address, on your own computer through the desktop app, or in the cloud with your key. Where it runs is shown on one page together with time and price.
  • Recipes instead of graphs: “Picture — fast”, “Picture — quality”, “Edit in words”, “Video with sound”, “Music”. A bot sees two or three fields, while the model, steps and sampler stay exactly as the recipe set them.
  • Editing what you send: attach a photo and say “remove the background” or “make it warmer” — the bot edits that picture instead of drawing a new one.
  • Your own graph becomes a recipe: export it from ComfyUI (Workflow → Export API) and the Studio finds the form fields itself, fixing everything else. From then on every bot can use it without seeing the graph.
  • Your GPU is not taken silently: the Studio runs one job at a time, yields the queue to your own ComfyUI tasks, says “not enough memory” before starting rather than forty seconds in, and always shows a Cancel button.
  • A long job goes to the background: a three-minute clip does not hold the bot’s turn — the finished file arrives as a card in the same conversation and as a notification in your channels.
  • Pictures end up where they are needed: the bot puts a cover, an icon or a background straight into the project files, and they ship to the subdomain with the site.
  • You can see where the time went: a library with recipe, station, seed and GPU seconds, a 30-day log, a daily job limit and the roles “Designer”, “Videographer” and “Musician”.
  • Folders in a project open up. The tree used to show three levels and stop at four hundred files — a file like src/components/ui/button.jsx simply did not exist; now any of them opens.
  • You can keep several files open: editor-style tabs with an unsaved marker. A draft edit survives a page reload.
  • Undo (⌘Z) no longer disappears when a bot writes a neighbouring file: the editor survives a full page redraw.
  • ⌘P finds a file by three letters of its name and also runs project commands — start, publish, run checks.
  • It got fast: the tree is read in one request instead of one per folder, size is computed in the background, and editing a file updates one row instead of the whole project.
  • File search understands regular expressions, case, “names only” and file type.
  • A project can now run several processes at once: front end, API and a worker — each with its own command, port and log. Previously the second start killed the first.
  • Projects got an environment: an API key is stored encrypted on the server and goes into the process at start instead of a .env file that travels to the repository. The bot sees a mask instead of the value.
  • Output streams, “npm install” no longer hits the command time limit, and a command can be stopped with Ctrl+C. Previous commands come back with the up arrow.
  • Preview became a tool: inner paths, four screen sizes, the page’s error console and the list of its requests. “Show the bot” sends it a screenshot together with the errors.
  • The preview page reloads itself as soon as a project file changes.
  • A bot’s changes are accepted hunk by hunk: take three out of five and send the rest back with a comment that becomes the bot’s task. A snapshot is taken before merging a branch.
  • You can see who wrote each line and compare any two versions, not just neighbouring ones.
  • The Overview tab answers “how is the project doing” on one screen: processes, address and its health, what the swarm did this week, what is waiting for you, size and free disk space.
  • A published address is checked every five minutes: if the site goes down you hear it in the project chat instead of from a client.
  • Dependencies show what is outdated, a database for the project is created with one button, and for your own domain the required DNS record is spelled out.
  • When there are many projects: one status screen, tags, an archive, search across all projects and moving a file from one project to another. It works from a phone too.
0.16.0

Canvas: a document you edit together

  • The Canvases section has been rebuilt: four modes — Read, Edit, Compare, History. The document looks the way the client will get it: headings, lists, tables and checkboxes instead of hashes and pipes.
  • Edits are saved as you type and no longer disappear: the draft survives a reload, ⌘S records a version.
  • If you and the bot edited at the same time, both edits are merged. Previously whoever came second got a “the canvas has changed” refusal and lost their text; now the disputed part is shown with both variants and you choose.
  • Any two versions can be compared, not just the last two. A corrected number is highlighted as a number, not as a whole paragraph.
  • The bot no longer edits silently: in a document you have touched it suggests edits and explains why. You accept or reject them, and the reason for a rejection reaches the bot — it will not suggest the same thing again.
  • Margin notes: select a fragment and write a remark — the bot reads it on its next turn and answers with an edit.
  • Select text and actions appear: rewrite, shorten, expand, translate, check the facts. Only the selected fragment and a map of the document go to the bot.
  • Documents can be brought in and taken out: Word, PDF, Markdown and plain text can be dropped into the section, and a canvas can be exported to Markdown, Word, a printable page or plain text. The bot prepares an email from the document as a draft.
  • Document kinds: text, code with highlighting, table, diagram (lines like “Request -> Review” are drawn as arrows) and a live canvas — a page written by the bot that runs in a sandbox.
  • Order at a hundred documents: folders, pinning, an archive and search across the text of all canvases.
  • It became cheaper inside: history is stored as differences rather than copies (31 versions of a brief take 5 KB instead of 1.2 MB), the list no longer carries document texts, and the bot gets a map instead of the whole document — 240 tokens instead of 9,700 for a forty-page brief.
0.15.0

Workshop: the app installs the bot’s tools by itself

  • The “My computers” tab now has a “Workshop”: Node, Python (via uv), code search, git for Windows, ffmpeg, pandoc and a browser — installed with one button, without Homebrew and without an admin password.
  • The bot’s commands on your computer now run with those tools. Before, the app used the system’s own paths: node and ffmpeg exist in your terminal, but the bot never saw them and answered “command not found”.
  • Tools are downloaded from our server and checked against a checksum; if it does not match, the install is cancelled. A tool counts as installed only after it actually ran.
  • A failed tool update rolls back: the previous version keeps working.
  • The bot can see what your computer has and ask to install what is missing — with your confirmation.
  • You can see how much space is used and clear the cache with one button. Everything lives in the app’s own folder: removing the workshop is removing one folder.
  • Long jobs on your computer: builds, npm install and tool installs now get up to half an hour instead of a minute, their output shows as it happens, and a “Stop” button ends them within a second and a half.
  • Projects on your own computer: the bot creates them from the same templates it uses on the server, under ~/VibeBot/projects, runs them on their own port and shows the log. Such a project runs on your machine only and is not exposed anywhere.
  • The bot can look at a page in your own browser — where the server-side one is helpless: personal accounts, sites you are signed in to. The screenshot comes back as an image, and you approve every address.
  • MCP connections can now run on your computer: a filesystem server sees your files, a mail server your mail. Pick the place in the connection card.
  • The app finds Ollama, Docker and LM Studio by itself and opens their installer on a button; next to it you see which model your particular machine can run.
  • The tools ship inside the installer: on a fresh machine the bot is ready even without internet. A “Remove the workshop” button deletes them completely and leaves your projects alone.
0.14.0

See where each model runs, and a “Find models” button

  • Every model shows where it runs: “server · GX10”, “computer · your laptop” or “cloud · api.z.ai”.
  • In the model picker the rows are grouped by place: your servers first, then your computers, then the cloud.
  • The “Find models” button discovers Ollama, LM Studio, vLLM, llama.cpp and Jan — on the server, on other servers in your network and on your computers.
  • Model lists refresh on their own; you can turn that off with the “search automatically” switch.
  • Models on your own computer are now available to bots: the request goes through the desktop app while it is open, and the data does not even reach the server.
  • The server name used in those labels is set on the “Models” page — with several servers you can tell which one a model is on.
  • Every cloud model now has an “Enter key” button: paste the key in place, the connection is tested right away and the model becomes available. Next to it, a “Get a key” link opens the provider’s key page.
0.13.0

Updates without reinstalling

  • When a new version reaches the server, an “Update” button lights up at the bottom of the window.
  • The button opens a “What’s new” window, so you see exactly what will change before updating.
  • The desktop app downloads the new version by itself and shows the progress; all that is left is “Restart and update”.
  • The Windows version now updates itself too: it downloads the archive, verifies the checksum and replaces itself on restart, keeping the previous copy next to it.
  • The website and the phone learn about a new interface right after it is deployed — no more reloading on a hunch.
  • If the window is hidden and nothing is typed in the message box, the interface updates by itself.
  • Settings → About now has a “Check for updates” button and the version history.
0.12.0

Projects, interface language and appearance

  • Projects: a bot builds an app on the swarm computer, shows it live and publishes it to a 24oros.ru subdomain.
  • The interface speaks Russian and English — switch it in the settings.
  • Eight themes, avatar sets for bots and motion styles.
  • Work on your own computer through the desktop app — only with your permission.
  • Owner rules, command risk assessment and command sandboxes.
  • Swarm mail: bots read the mailbox and prepare draft replies.
Next