Cerevisor 3.8.0: What's New

3.8 gives every agent a conversation that lasts. Each agent card now has one chat that survives a restart, a Stop and new runs, and every message you send goes through the same checks a run does. Around it: Cerevisor connects the AI tools already on your computer at first launch, Claude Code and Codex agents can use your connected tools, the run report opens by itself when a run ends, and a shared workflow can be added to the world you have open.

1. A lasting conversation with each agent

Each agent card has one conversation, kept on this computer. It is still there after a restart, after a Stop, and after later runs of the workflow. You can talk to an agent on every kind of provider, including Claude Code, Codex, Cursor, Antigravity and Grok Build.

A chat message is held to the rules of a run. The agent answers with the tools, permissions and instructions on its card. Its tool calls go through the same permission checks, approval setting and tool hooks as a run, and each reply is recorded and listed in the System Log as a Chat turn. A message to an agent that has finished now continues the conversation. Before, it ran the agent again.

Easier to use. Replies stream as formatted text with a card for each tool call. The box grows as you type and is never locked. Paste, drop or pick any file. A message sent while the agent is replying reaches Cerevisor’s own models and Claude Code at once, and waits in a queue on Codex, Cursor, Grok and Antigravity, where a queued message can be edited or sent with Send now. Stop is there for the whole of a reply. Reset starts a clean conversation, and Earlier conversations keeps the old ones.

The chat follows the run. Every agent that finishes in a run gets a line for that run in its conversation, so the chat knows what the agent just did. An agent’s chat opens on Activity, the live log of what the agent is doing, with a Chat tab beside it and the same box under both. Sending a message shows Chat; a new run brings Activity back.

It says what it can do right now. One status line reads Replying, Steering, Waiting for its turn or Working in the run. A run keeps the models it started with, so when you change an agent’s model while its workflow runs, one notice says the change applies from the next run, and the chat goes by the model the run is really using.

A conversation that read content from outside (a web page, for example) stays more careful with what it sends out until you press Reset. What agent chats cost is shown in Analytics, under Total Cost.

2. Connect the AI already on this computer

On a computer that has never had a provider in Cerevisor, the first launch looks for Claude Code, Codex, Grok Build and Ollama. Each one that is installed and signed in is connected without a click, and one notice names it, with Review.

In Your model providers, a tool that was found and is not connected shows Found here and a Connect button: one press, nothing to fill in. This is how Antigravity connects. A connected tool reads Connected with a small Manage, and Manage no longer tests the connection by itself; the test is a button.

One sentence at the top of the column replaces the box you had to tick for each tool: “These apps set their own permissions.” It means what it says. A tool like Codex or Claude Code runs under its own rules, and Cerevisor tells you so once, in one place.

This happens once per computer. After anything has been connected, by you or by Cerevisor, no later launch connects a tool by itself.

3. Connected tools: one way to add them, and more agents can use them

Add MCP server is one box. Paste a URL, a config or a command, and Cerevisor works out which it was given. Popular servers are Suggested cards under the box. A remote server has an Authentication setting: sign in with your browser, a bearer token, or none, with custom headers under Advanced.

Tokens are stored sealed. A token or header value you type is kept in your operating system’s secure storage, never in the settings file. It is used only by its own server at the address it was saved for. Change the address and the old token is not sent, on Test or on Save.

A sign-in that cannot work says so. GitHub’s server does not let an app sign in with the browser the way Cerevisor does, so its card now asks for a token and tells you where to make one. Any other service that does not offer browser sign-in answers with one sentence and Use a bearer token instead, where before it showed a technical error.

Claude Code and Codex agents can use your connected tools. Before, these agents were shown tools they were never handed. Now each call they make to a connected tool is made by Cerevisor, through the same checks as any other tool call. Cursor, Antigravity and Grok Build agents cannot receive connected tools yet, and Cerevisor says so in the agent’s settings, in the summary before a run and in the run window, instead of listing tools the agent does not have.

Because Cerevisor cannot see everything such an agent reads, it is careful with remote servers you have not marked Trust this server: a call to one asks you first when you are watching, and is refused in a Hands-free or scheduled run.

4. The run report opens when the run ends

When a run you started in the window finishes, its run report opens in Cerevisor by itself. If something is covering the canvas, a notice says The run report is ready. with Open report.

• Links work. A link in the report to a file the run produced opens that file in the File Viewer, with a Back button to return. A link that points outside the run’s files opens nothing and says so. Web links open in your browser.
• Plain technical English. Agents are asked to write the report in ASD-STE100, a standard for short, clear technical sentences, with a link for each file.

Runs started by a schedule, the Operator, a script or your phone still write the report; they do not open it.

5. Add a shared workflow to the world you have open

The share menu at the top right has Import shared workflow…. Pick a shared file, and its workflows are added to the open world as new workflows, with their files in a new folder. A file that holds several workflows shows a list to tick. Each added workflow is one step you can undo.

Every rule a shared file is held to is unchanged: nothing of the sender’s computer comes with it, and none of its approvals travel. Import from Clipboard is removed.

6. A help assistant that sees what you see

Need a hand? now knows where you are. With each question it gets a short description of your screen: the open world, the workflow in focus, the selected agent, and that workflow’s last run and error. It reads the help pages one section at a time, so no help text is out of reach, and it can explain a run, list a workflow’s problems, describe an agent and explain how data moves between agents.

It still cannot build, edit, save or run anything. The screen details go to your default model provider with each question, and the help page says exactly what is sent.

7. Operator tells you what it built

When Operator builds or improves a workflow, a notice on the canvas says so, with Show, and each workflow it built has a card under Produced with Show on canvas. A later check keeps improving the copy Operator already made, instead of making another copy. A workspace summary is now a note to read, with Dismiss, where before its Accept button could only fail.

8. A switch for where agents may work

Settings → Admin → Keep agents inside the workflow’s folder is on by default, as before. Turn it off, after a confirmation, and an agent’s file tools can read and write anywhere on this computer, except places no switch opens: folders that hold keys and sign-ins, network shares, and Cerevisor’s own data, which agents cannot change. The switch applies from the next tool call, also in a run already going.

9. Smaller things

• The working folder is one click away. A folder button in the workflow controls names the folder the workflow works in and opens it.
• The File Viewer shows when a file was created and last modified.
• One message box before a run. The launcher is now a drop zone and one large box. Paste a screenshot or a copied file with Ctrl+V and it is attached.
• File names in the chat builder are links. A file the builder mentions opens in Cerevisor when it really exists.
• Builder answers stay with their question. When the chat builder asks several questions at once, each answer sits under its own question.
• Model search. The Default model list in a provider’s setup can be searched, and a long price list folds away behind Show prices.
• Settings is easier to find. The gear is named Settings, Ctrl+, (Cmd+, on a Mac) opens it from anywhere, and macOS has Cerevisor → Settings… in the menu bar.
• A secret from where you need it. The vault picker on an agent has New secret, and an empty vault says so.
• Install and update. The Windows installer has a Welcome page and a status line that follows the real install. An update reopens Cerevisor by itself. The launch screen follows your theme.
• Memory. Proposals group into themes again once you have finished reviewing the last group. The Profile facets panel no longer closes when you type in it.

Changes to how every run behaves

• A condition is checked against what its connection hands over. With a File Reference handoff, that is the files the agent before it wrote during this run, read in full, with its reply. A file left by an earlier run is never used. When conditions skip every final agent, the run ends as Finished, no result, not Complete.
• A model changed while a workflow runs applies from the next run. The run keeps the models it started with.
• Chat counts. A chat reply counts as the workflow running, and a workflow that gained new powers shows the summary before a chat message is sent. The Agent-chat cap setting is gone. Chats from before this version are not carried over.
• A workflow with a Claude Code or Codex agent and an enabled connected tool asks once to confirm that the agent can now use connected tools. Scheduled and other unattended runs of that workflow wait for that confirmation.
• The launcher no longer has suggestions or a Smart Mode picker. A run started from it uses the workflow’s own settings.
• The browser tool checks local files. A file: address outside the workflow’s folder is refused unless the file is declared on the agent, also when a page tries to go there by itself, and a page can no longer open a new window.
• Erasing run history also erases agent conversations. The reset category is now named Run history, analytics & agent conversations.
• Codex. Cerevisor now runs the same Codex a terminal runs when more than one copy is installed. Update Codex says what really happened to that copy, and a model that Codex refuses because it is too old is explained in plain words. The check before a run names a damaged Codex sandbox file, and shows one card per cause instead of one per agent.

Also fixed

• An agent whose role this version does not know took down the whole window. It now opens as Custom.
• After adding two workflows and undoing one while looking at another, the canvas and the world could disagree about which workflow was active. Undo now keeps them together.
• On a Mac, the green full-screen button no longer hides the title bar.
• The line under the Activity and Chat tabs is no longer covered by the approval pills.
• A handoff by file reference passes on the agent’s own words and accepts a path with spaces.

Security

• A few reported weaknesses remain in a library that the Cursor integration brings in. Cerevisor does not use the affected parts. They go when Cursor’s SDK moves to the fixed version.
• Tokens for remote tools are sealed in your operating system’s secure storage and are tied to the address they were saved for.
• Chat is held to the run’s rules. The tool calls of a chat message go through the same checks as the tool calls of a run.
• The folder switch has a floor. With the limit off, agents’ file tools still cannot reach folders that hold keys or sign-ins or network shares, and cannot change Cerevisor’s own data. A shell command is checked only by the words it contains.
• A page shown in the File Viewer can open only files of its own run.
• Checksums: the SHA-256 hash of every installer is published next to it on the download server.
• Windows installers are not yet code-signed, and macOS builds are not yet notarized.

Known limits

• Agent chat: an agent that is working in a run on Codex, Cursor, Grok or Antigravity cannot be given a message until it is done. A model changed during a run does not reach that run, also for an agent whose turn has not come.
• Connected tools: Cursor, Antigravity and Grok Build agents cannot use them yet, and neither can a Claude Code credential saved in Plan mode. Codex was tried on version 0.160 only. A Claude Code or Codex agent still loads the tools you set up in that app yourself, and those calls are not Cerevisor’s to check.
• Browser sign-in for connected tools works only with services that let an app register itself. GitHub and Slack do not; GitHub’s card asks for a token, and Slack’s suggested card still offers a sign-in that ends in the sentence above.
• Finding your tools: Claude Code on a Mac is not found at first launch, because its sign-in is kept in the Keychain; sign in from the provider screen as before. A tool is connected without a test, so an expired sign-in shows on the first run.
• The run report is always in English. Plain technical English is asked for and not checked. While you are clicking inside the report, Esc does not close the viewer; the close button does.
• The help assistant sends more text with each question than before, so a question costs a little more, and a local model with a small context window may not have room for it.
• The folder switch: Codex in its default sandbox still cannot write outside the folder, and Claude Code and Antigravity agents follow their own rules whatever the switch says.
• Factory mode: Cerevisor saves an open workflow by itself. For a workflow file that was written by hand or last saved by an older version, that save can land while an item runs, and the factory then pauses once with “The workflow was changed while the factory worked.” Press Start again.
• Everything listed under 3.1.0’s through 3.7.1’s known limits still applies.

Platforms

Available for Windows, macOS (Intel and Apple silicon) and Linux (AppImage and Debian package).

Updating

Existing installations receive the update through Cerevisor’s built-in updater, and the installers are available from cerevisor.com/download. Your workflows, run history and settings carry forward untouched.