Cerevisor 4.1.0: Eight languages, sharing with everything it needs, an Operator per world

Cerevisor now speaks eight languages. You can share one agent or a whole workflow with its connected tools, skills and sealed secrets, and lock it against editing. Each world gets its own Operator, and the Overmind is one assistant you can chat with above all of them. A group on the free canvas can have a lead that splits the work. Stop now always stops, and a run marked needs attention says what needs it.

1. Cerevisor in eight languages

The app, Cerevisor’s messages and the phone companion are now in English, Russian, French, Spanish (Spain), Portuguese (Portugal), Simplified Chinese, German and Italian. Choose one in Settings → Language. The change applies after Restart now, which saves the open world first. On first start Cerevisor uses your computer’s language when it is one of these, and English otherwise. Austrian and Swiss German open in German, Swiss Italian in Italian. The phone follows the phone’s own language.

• What stays English: product names, text only a model reads, what is saved into your files, copied logs, and the help pages.
• Dates and numbers in an English app follow your computer’s region, as before.
• How it was translated. The translations were made by a model and reviewed, with permission prompts and confirmations of actions that cannot be undone checked line by line. No native speaker has read them yet.

2. Share an agent or a workflow with everything it needs

An agent card’s right-click menu has Share agent…. A shared file can now carry the agent’s or the workflow’s connected tools, fully set up, its skills, and, if you choose, its secret values.

• Secrets are off by default. Turned on, they are sealed with a code of at least 8 characters, which you pass on separately; Suggest a code makes one. Case, spaces and hyphens in the code do not matter.
• Receiving. The import dialog lists what came with the file. A shared agent lands in the workflow that is open, with Add to this workflow. Sealed values need the code, or choose Import without the secrets. Secrets go into your Vault for that workflow. A name already taken is stored as <NAME>-2, and your own entry is never touched. Anything that still needs your own value is listed under Still needed. A skill of the same name that you already have is kept, never overwritten. If any step of an import fails, everything already written is undone.
• Installing the extras needs a paid plan (the trial, Pro or Power). Import without them brings the bare agent or workflow on any plan, and sending is free on every plan.
• Lock editing with a code. The shared copy shows Locked in the footer, with Unlock to edit for this session and Remove lock for good. A locked workflow can still be run, rearranged and given your own provider, model and Vault secrets. Loop & Enhance, Endless and Progressive refuse a locked workflow before anything is spent, and the Operator may run it but not change it. The lock guards against changes made by mistake. It is not a security wall: the file is plain text.

3. One Operator per world, and the Overmind

Each world now has its own Operator. The Operator panel’s new Worlds tab lists the worlds Cerevisor knows. Each has an Active or Test switch, its own checks switch and an optional daily limit. Checks take turns, the longest-overdue Active world first, and the footer names the world being checked. Each world keeps its own history and report. The Report has a world picker, and the history from before this version is kept read-only under Before worlds.

• A world that is not open can still move forward. An Operator can advance, build and improve a world that is open in no window. The change is saved into the world’s file, and refused if the file changed since it was read. It has no undo on a canvas.
• A copy is its own world. After a Save As to a new file, the copy is kept as its own world and says so.
• A world open in any window cannot be forgotten. Forget never deletes the file.

The Overmind is one assistant above every world. Press Overmind in the Operator panel: the worlds are on the left, the conversation on the right. It reads a summary of every world and opens one world or workflow at a time when it needs more. It acts in two ways only: it files a goal in a world, or it suggests a line for your memory. A filed goal waits for your Approve unless Let the Overmind start work is on.

• Spending. At most 6 look-ups per message. Its spend counts in the same daily total as the Operators’. If a cost cannot be priced, Cerevisor stops Operator and Overmind spending until the next day, and says so.
• Check on a timer has its own switch, and Every takes an interval you set, from 30 minutes to 24 hours. Both switches start off, and turning one on needs a plan that covers the Operator.
• Chatting with the Overmind is open on every plan.

4. Groups that work together

A group on the free canvas has two new switches. Introductions asks each agent to introduce itself on the board and work with the others. Lead adds a lead card to the group. Your message then goes to the lead alone, which splits the work, hands each agent its part by name, checks what comes back and replies with one answer.

• Files and pictures in a group message: the paperclip, a paste or a drop. With a lead, each agent gets the files once, with its first assignment.
• The group box shows a face for each agent (the lead with a small crown, a still green dot for one that is replying) and grows with its text up to six lines.
• Every agent gets a distinct name. New, pasted and duplicated cards are numbered (Developer Agent 2, … (copy) 2), in waves too. Turning Lead on numbers the group’s agents that share a name, and one Undo puts them back.
• A note to a name two agents share is refused and nothing is sent, where before it went to the first card of that name.
• A lead’s note to “anyone” starts every agent of its group, unless the note only reports work that is done.

5. Stop always stops

An agent that hung could keep a run going even after Stop. That can no longer happen:

• Stop ends any agent within about 25 seconds, on every provider, Cerevisor’s own agents included.
• A Codex run ends when Codex ends, even when a program it started still holds its output open.
• Programs an agent started are tracked. After a run that ended badly (Stop, a stall, a usage limit) they are stopped. After a normal end they are left running, since an agent may start a server on purpose, and the System Log names them.
• Codex’s usage limit ends the agent at once, with the time Codex says the limit resets.
• The chat says what happened. The header says the agent hit a problem, or that it is stopping, the error shows as a red row, and a message waiting to go reads Not sent, with Send now.

6. “Needs attention” says what

A run that left something for you used to show only an amber mark. The world’s hover card, the workflow rail’s tooltip (Left for you:), a workflow’s hover card and the footer’s run dot now list what the last run left: changes not committed, an output not written, a check that failed, what an agent could not check. A run from an older version says Open its last run to see what was left for you. An agent that was skipped now has an amber output card, not a green one.

7. Human steps: a tab for each waiting step

• A tab for every waiting step, oldest first, each with Waiting since … and its workflow. A notification and a card’s Open now show their own step, not the oldest one. What you typed in one tab survives a look at another.
• Markdown is laid out. Tables, lists and code in agent reports show properly on the plain form and on pages the AI Builder writes. Pages written before this version show raw text until you ask the AI Builder to write them again.
• A page whose script cannot run is refused when the AI Builder saves it, and a run shows the plain form with the reason instead of an empty page. Long pages written by an Antigravity builder no longer arrive broken.

8. Skills from tips

• Make skill and equip is on every tip a skill can come from: a repeated tool failure, a reliable recipe, the same answer to the same question. It makes the skill and puts it on the agent, as one Undo step.
• Settings → Automation → Equip agents with the skills made from tips does it after runs on its own. It is off by default, and nobody reviews these skills before they reach the agent’s instructions.
• The “reliable recipe” tip no longer fires on agents that only open, list and search files.
• The unused-skill tip now sees Cerevisor’s own skill calls, Claude Code’s skill calls and reads of a skill’s file. Antigravity, Codex, Cursor and Grok Build agents no longer get it, because Cerevisor cannot see their skill calls. A tip that no longer holds goes away after the workflow’s next run.

9. Analytics counts more of what you spend

• The help chat’s spend now shows in Total Cost, Spend over time and every token figure, with Includes $… from the help chat and a Help chat row. A help question is never counted as a run. Questions asked before this version are not counted.
• List prices for Antigravity and the Chat Builder. With subscription runs shown at list price, Antigravity runs and new Chat Builder sessions now get a figure. Prices come from OpenRouter’s public price list first, then the bundled tables. A model on a local endpoint is never priced.

10. Smaller things

• Groups of similar suggestions. On Memory → Suggestions, near-duplicates are grouped into one card that says how many it combines, with Show to answer them one by one. Accept saves one line for all of them. Reject rejects them all, so none comes back.
• Install Codex works without Node.js. On a computer without npm, Cerevisor runs OpenAI’s own installer, and on Windows only when OpenAI’s signature on it checks out. A missing npm is now found in any Windows language.
• Restarting for an update while workflows run asks first, with Restart anyway and Later.
• The provider check before a run checks the sign-ins your agents really use and names the one that failed.
• Antigravity agents are told they have their Vault secrets. They always received them, but were told they had not and stopped with a blocker. Cursor on this computer is now the only provider that cannot receive Vault secrets.
• Show all steps on a sub-agent card in a Claude Code chat shows every step the sub-agent took.
• Progressive puts a new agent’s files in the workflow’s own folders, repairs inputs that point to the wrong folder, and one step that could not start no longer blocks every later step.
• Show on canvas moves the camera to the workflow.
• Settings → Subscription → Manage or cancel subscription opens your billing page, already signed in.
• The tour on your own task leads with the builder’s messages, each with Skip tour.
• On Windows, a program that ends by itself no longer leaves a hidden console window running in the background.

Changes to how every run behaves

• Agents’ work gives way to your apps. A new switch in Settings → Admin, on by default, starts Claude Code, Codex, Grok Build, Copilot, Antigravity and Cerevisor’s own agent commands below normal priority, so your other apps stay responsive. A delegated agent may answer more slowly on a busy computer. Cursor is not covered.
• Stop ends an agent that has not stopped after about 25 seconds, on every provider.
• A run that starts part-way checks only the agents it runs. A missing input of an agent the run skips no longer refuses the run.
• New agent cards get numbered names when a name is taken.

Security

• Electron 44.10.0 (Chromium 152.0.7977.169, Node 24.22.1), unchanged from 4.0.0.
• Two fixes that matter for your safety:
• A long Antigravity prompt is private to your account. A prompt too long for the command line was written to one shared temporary folder that, on Linux, every account on the computer could read. Each prompt now gets its own private folder, removed after the run. Windows was not exposed.
• Antigravity agents are told the truth about their Vault secrets (see Smaller things), and the check before a run and the help page now agree with it.
• What a share can carry. Secret values travel only if you turn them on, sealed with your code. A received secret named like a coding tool’s sign-in key (OPENAI_API_KEY, GITHUB_TOKEN and the like) never reaches Claude Code, Codex, Cursor, Copilot, Grok Build or Antigravity. A received connected tool that came with every value it needs starts when you press Import, so read what a file brings before you import it.
• One new request: opening Analytics fetches OpenRouter’s public price list, at most once a day, with no key, no account and no run data. It is the same request the model catalogue’s Refresh already made.
• Windows installers are not yet code-signed, and macOS builds are not yet notarized.

Known limits

• Translations are model-made and have had no native review. On an iPhone the companion app may stay in English.
• Sharing has not yet been tried between two computers.
• Groups with a lead, the Overmind and world Operators are new: how well a lead splits work and how well the Overmind decides will improve with use.
• Install Codex without Node.js has not been tried on a Mac or Linux.
• Show all steps works only for Claude Code chats made from this version on.
• List prices: older Chat Builder sessions stay unpriced, and an Antigravity run’s list price counts every input token at the full rate.
• Needs attention: most runs will still be amber, because any item an agent lists as Not verified counts. The mark stays until the workflow runs again.
• Nothing in this release was tried on a Mac. It was built and checked on Windows.
• Everything listed under 3.1.0’s through 4.0.0’s known limits still applies.

Platforms

Available for Windows, macOS (Intel and Apple silicon) and Linux (AppImage and Debian package). Using apps is on Windows and on a Mac only; on Linux the setting does not appear.

Updating

On Windows and Linux, existing installations receive the update through Cerevisor’s built-in updater. On a Mac, download the new version from cerevisor.com/download and replace the app in Applications. Your workflows, run history and settings carry forward untouched.