Cerevisor 3.7.1: What's New

3.7 brings Factory mode: write down what you want built in a project, and Cerevisor works through the list one item at a time. It runs your workflow on each item, has an independent checker judge the result, tries again when the check fails, and commits each finished item to its own local branch. You can stay at the desk and answer its questions, or leave it to run. Around it: programs on your computer can now run your workflows, approvals show you what you are approving and let you ask for changes, and new models reach every picker the day they ship.

1. Factory mode

Factory mode turns one of your workflows into a worker for one project. You queue the things you want built, one item each, in the order you want them. Cerevisor runs the workflow once per item, one at a time. Each item gets its own folder of working files inside the project and its own run, which you can watch on the canvas like any other. In a project that uses git, each try is committed on a branch called cerevisor-factory, so the history shows how every item got there. Nothing is ever pushed. Cerevisor never pushes, merges, rebases or opens a pull request; what happens to the branch is up to you.

Start in one click. Deploy the factory, in the workflow’s controls beside Schedule, opens Settings → Factory mode. From there one press lays out a ready-made factory of seven agents and switches Factory mode on, as one step you can undo. The ready-made factory researches the request, agrees an approach with you, writes a plan, builds, and then checks the build with an agent that had no part in it.

Done means checked. The planner writes down what “done” means for the item. A separate validator judges the work against that and against your request, and writes its verdict. The verdict only counts if the validator did not change your project files to reach it. An item is done when the check passes, not when the builder says so.

Tried again until it works. A failed item is tried again, up to three times in all by default (you can choose one to five). Each new try starts from the last one’s work and is told what went wrong. You can also set a spending limit per item, checked between tries. When nobody is at the desk and you set no limit, an item stops at US$5.

With you, or without you. - Attended: the factory asks you when it needs to. Approvals and questions wait for you, with no timer. - Unattended: questions are decided by the model and approvals pass at once, and every choice is written to the item’s decisions.md, so you can read later what was decided and why. Unattended needs your confirmation on the workflow and Hands-free switched on on this computer. Agents whose only job is to ask you, such as the ready-made factory’s approach agent, are left out of an unattended item.

Before you start, the factory finds the project’s own test, build and lint commands and shows them in one line. Check the commands work runs the test and build commands once, so you know whether anything will check what the items build. Each item can carry pictures (paste, drop or pick, up to 10 per item), and every agent in the item whose model can see pictures gets them. Other programs can add items to the queue too (section 2).

The factory panel shows each item and how it is going. A row waiting for you says so, and the review panel (section 3) opens by itself when an attended item needs an answer. Smart routing (section 5) is on for every item: it picks a model for each Auto agent on every try, and a try after a failure asks for a stronger one.

Factory mode is part of Founding membership and is included in the 7-day trial.

2. Run workflows from scripts and other programs

Turn on Settings → Automation → Local API, and a script, a terminal command or another program on your computer can run any of your saved workflows. It can follow the run, answer its questions and read what it produced. - Programs that understand MCP, such as Claude Code or Cursor, connect with Copy MCP config. - For scripts there is a command: cerevisor run <workflow file> --wait --out <result file>, with cerevisor answer and cerevisor status beside it. - cerevisor factory add adds an item to a project’s Factory mode queue from a script or a build server. It adds work; it never starts the factory.

The local API listens only on your own computer and is off until you turn it on. Every program needs a secret token, which you can replace at any time with Rotate token. A run started this way asks its questions exactly as a Run button run does, in the app window, on your phone or from the command line, and nothing is approved for you. The local API is part of Founding membership and is included in the 7-day trial.

3. Approvals you can read, and Ask for changes

Every run now has one review panel for its approvals, questions and permission requests. An approval shows the documents you are approving, up to four of them, right in the panel. You have three answers: Approve, Stop here, or Ask for changes with a note. Asking for changes sends the agent that wrote the work back to revise it with your note, up to three rounds per approval. If the work changed since you last looked, Cerevisor says so and asks again before your approval counts. The next agent is told that you approved the work, and when.

Requests waiting for you also appear in a banner under the top of the window; click one and the canvas flies to the agent that is asking.

4. New models without waiting for an update

Each saved provider in Your model providers has Refresh models, and the header has Refresh all models. Cerevisor asks the provider which models your account can use today, and new ones appear in every picker at once, without a restart. You can also refresh from inside an agent’s model picker. Where it can, Cerevisor fills in each model’s price, how much it reads at once, its reasoning levels and whether it takes pictures. - Update Codex, Update Claude Code and Update Grok keep those tools current from the same screen. Claude Code is downloaded into Cerevisor’s own folder, checked against the publisher’s fingerprint, and used only after it starts cleanly. - A model with no known price still runs. When a spending limit is set, a red line warns “Your spending limit can’t see this model’s cost.” before the run and in the run window, and the cost line names the model whose usage it cannot count. - With your own Artificial Analysis key, benchmark scores help place a new model. - A failed or empty refresh never replaces a list that worked, and a refresh never changes a model you picked.

5. Smart routing for Auto agents

Turn on Auto-route agents to best model in Settings → Integrations → Smart Routing, and before the run each agent set to Auto gets a model on its own account: a quick task gets a quick model, a hard one a stronger model, and never a model from another provider. With TypeSafe Jev connected and Smart routing ticked under its checks, Jev decides; otherwise your background model does. Settings has a slider for how sure Jev must be, for each of its checks. Codex and Claude Code agents are routed for the first time. The run window says which model each agent got, who decided, how sure it was, and why a default was kept.

6. The output check is your choice

When an agent finishes, Cerevisor can ask a model whether the agent did what it was asked, and give it one try to fix it if not. Each check costs a model call, so it is now your choice, and it starts off. Turn it on in Settings → Integrations → Output check → Check each agent’s result. The same switch covers the check that decides whether an agent saying it is blocked really is. For work that matters, a reviewer card is the stronger check.

7. Operator builds while you are away

Operator can now build a new workflow for one of your goals, improve an existing one on a branch, or add a skill to your skills. The work happens in the background, lands on the canvas where you can see it, and is linked to the goal it serves. Operator’s own builds and branches hit the same limits a build of yours does, and its journal says what it built. Operator is part of Founding membership and is experimental.

8. A help assistant, and an easier start

• Need a hand?, at the bottom-right of the window, opens a help chat. It answers from Cerevisor’s help pages and can look at your own worlds, workflows and last runs. It never changes anything; when you ask for a change, it tells you how to make it.
• A new start for new users: two buttons, Guide me through my first workflow and Explore on my own, and nothing else. The guided first project includes a short tour of the first agent and the chat builder.
• Simple view, the eye button in the title bar, hides the extra panels around the canvas. Switching between simple and full view moves and resizes every part into place in under a second, and is instant with reduced motion on.
• When a run finishes, its output panel comes into view beside the Workflow Builder and the files it produced flash once.
• A map of the workflow sits in the canvas corner; drag it to move around a large workflow.
• Grok Build installs in one click from the Provider Hub.
• Memory advisories now live in Analytics instead of a banner over the canvas. The Settings tabs work with the arrow keys.

Changes to how every run behaves

• New workflows start with web access on, able to reach any site, so search and fetch work out of the box. Workflows you already saved keep their own setting. The protection against prompt injection still applies: after an agent reads untrusted content, sending anything out asks first, or is refused when nobody is watching.
• An agent whose every source was skipped or failed is now skipped, instead of running on nothing. The same goes for an agent whose input comes from an agent that was skipped or failed in this run; before, it only warned. A sole downstream agent of a failed agent set to Skip & Continue no longer runs.
• A file can be kept in each run’s own folder. Tick Keep in each run’s own folder on a file reference, and each run writes and reads its own copy. If no agent writes that file, Cerevisor mentions it before the run instead of refusing to start.
• The output check, when on, reads each file on its own allowance, and when it cannot judge a file it names the file and the reason.
• Decision cards can answer that none of their outcomes fits, and their results are easier to read.
• Spending caps: a hosted OpenAI-compatible account (api.openai.com, OpenRouter, xAI) whose default model has a known price now stops at the US$5 default cap, where it was uncapped. After a refresh, OpenRouter- and xAI-hosted models are billed at their listed price, where they counted as US$0 with a warning.
• The chat builder’s default Antigravity model is now Gemini 3.8 Flash.

Also fixed

• A workflow whose agent had no saved position on the canvas crashed the window when opened. It now opens and places the agent.
• The Add MCP server and Add vault entry dialogs now cover the window instead of being cut off.
• An error in one part of the window now shows a message there instead of a blank app.

Security

• Updated to Electron 44.4.5, which brings the current browser security fixes.
• Advisories fixed in the updater and in several bundled libraries.
• A few reported weaknesses remain in one part of a third-party library that Cerevisor never uses. They go when that library’s maker releases a fix.
• Checksums: the SHA-256 hash of every installer is published next to it on the download server, so you can check that the file you downloaded is the one released.
• The File Viewer opens more: ordinary files anywhere on your computer’s own disks. It still refuses folders that hold keys and sign-ins, network shares and programs, whatever you have granted.
• The local API (section 2) is a new way in from programs on your computer: only on this computer, off until you turn it on, and protected by a token.
• Windows installers are not yet code-signed, and macOS builds are not yet notarized.

Known limits

• Factory mode: no phone view of the queue yet, and no link from an item to its run view. Build output that is not in .gitignore, and that the checker’s own test run rewrites, counts as the checker changing your files, so put build output in .gitignore. Pinning a board note on a factory workflow while it works pauses the factory. Choosing models for a try usually costs one small extra model request. Pictures cannot be added through the local API or the command line.
• Local API on Windows: Cerevisor.exe is a windowed program, so PowerShell and cmd do not wait for it. Scripts use Start-Process -Wait -PassThru or start "" /wait; the help page shows how.
• Copy MCP config can issue a new token, so programs set up earlier need the new config.
• Model details: Gemini, Anthropic and Cursor listings carry model names only, not their limits or features. Refreshing Claude Code’s models starts Claude Code briefly, with no message sent.
• Smart routing does not route Antigravity, OpenRouter or Ollama agents. A TypeSafe Jev connection made before 3.7 has Smart routing off until you tick it.
• Simple view: with about ten full agent cards on screen the switch is less smooth, and an open menu can be cut off for a fraction of a second.
• Everything listed under 3.1.0’s through 3.6.0’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.