Cerevisor 3.3.0: What's New

3.3.0 is mostly about the first hour. Connecting an AI model now starts by asking what you already pay for and never asks you to open a terminal; a guide walks you from a blank app to a first finished result; the small yes-or-no judgments Cerevisor makes while a workflow runs are visible for the first time, and can be made near-instant with an account of your own; and the app has five color schemes instead of two. It also adds a second, cheaper plan, and makes two experimental features part of the top plan. If you are an existing user, the one thing to read is this: the team conversation, where your agents leave each other notes during a run, is now part of Founding membership. It has been available on the free plan since 3.1.0, and from this version it is not - but if you already have a license you are a Founding member, and you lose nothing. Every license sold before this release covered everything, so it still does; update the app and the Subscription tab will say so.

1. Connecting an AI model

The setup pane now opens with a question you can answer: What do you already have? - ChatGPT Plus or Pro, Claude Pro or Max, A Google account, An API key, or Nothing yet - run free on this computer - and takes you straight to the right place. Tiles lead with the plan you already pay for rather than the tool’s product name, and each one says what you need and roughly how long it takes before you commit. The tiles your answer can point at are shown first; the rest wait behind More options.

Nothing asks you to open a terminal. Codex and Antigravity install with one click - Cerevisor runs the install, shows exactly what it runs, and tells you plainly how it went, including the awkward case where the install worked but the tool is not findable yet. Claude Code needs no install at all: Cerevisor carries its own copy and uses yours when yours is newer, and signing in no longer needs a token pasted from a terminal.

Sign-in waits end. Help appears after a minute, the wait stops after five and says so instead of spinning, and Cancel works at any point. The Antigravity panel finishes on its own once you come back from the Google sign-in.

A connection check has to prove itself. A test passes only when the tool exited cleanly and gave a real answer; one that times out, comes back signed out or returns nothing is now reported as a failure rather than quietly passing. While something is missing, Save is greyed and says why in one line - Install it first., Sign in first., Checking your connection… - and if the check fails it becomes Save anyway, so you keep the escape hatch. When the first connection saves, a banner reads You’re connected and offers Build my first workflow.

Less on screen when there is nothing to choose. With one connection you see just Default model; a background model, a builder model, what to fall back to at a limit, a spending cap and reasoning effort sit under More model settings, which appears once there is a second connection to choose between. Codex’s screen is now two steps, Install and Sign in, with everything else under Advanced. And a Windows account whose folder name contains a space no longer reads as signed out forever; provider tools inherit your own proxy and certificate settings; and a Codex installed after Cerevisor was launched is found without a restart.

2. Your first project

A brand-new installation now opens a guide instead of an empty canvas: Let’s finish your first project. Four steps - Your project, Connect AI, Review and run, Your result - and the promise in the title is the point: you leave with something you can use, not a tour.

You start by choosing Try a guided example or Start with your own task. The examples are small and honestly fictional - Notes into an action plan, An idea into a project brief, Details into a polished email - and every one of them is editable before you run it, so you can make the sample yours in the same box. If you bring your own task, your brief is handed to the builder and a team is shaped with you. Beside it, a panel shows how a workflow actually works: your brief, one agent per job, the connection that passes the work forward, and the kind of result to expect.

The guide then gets out of the way. Once your project exists it shrinks to a small companion in the corner that follows the real run - each agent with its own status - and its button changes to whatever is next: Run my first workflow, Follow the live run, Open your result. At the end it shows what your team made, with Copy result and one closing lesson for your next project. Explore on my own dismisses it at any point, Hide guide tucks it away, and nothing about it is one-way: Guide me through my first project and Reset guide on the Home screen start it again, and your saved projects, runs and connections stay intact. Existing installations are never pulled into it.

3. Quick checks

While a workflow runs, Cerevisor makes a few small judgments: is this connection’s condition met, should this loop stop, did this agent do what it was asked, did this run go well. Those judgments now work better, and you can finally see them.

Fixes that need nothing from you. If your workflow’s default was a tool that runs its own work - Codex, Cursor, Claude Code, Antigravity, Grok Build - every conditional connection used to fail its check and skip the agent after it, even when the condition was plainly true; Cerevisor now finds a model that can answer, and if none can, the run says why instead of silently skipping. A check that succeeded used to leave no trace at all, so a finished run could not say why one branch ran and another did not: a checked connection now carries a chip on the line (yes · 1.4 s) and the run window shows one line beside the cost (23 quick checks · 2.1 s · $0.0003), both rebuilt from the run’s own recording when you reopen it later. Checks used to read only the first part of a long output, so a verdict in the last paragraph was invisible - they now read the beginning and the end, with secrets stripped first. A loop-exit check that comes back unsure ends the loop instead of spinning another round.

Decisions you can put on the canvas. A decision node asks one bounded question and routes the work by its answer, and every one of them has an Unsure path you can wire to a regular agent. Click it to set the question, a Decision rubric saying what each outcome means, and, for document-dependent decisions, Require source-matched evidence: an upstream agent supplies the source file and the verbatim passage, Cerevisor checks the passage really occurs in that file before asking anything, and missing or changed evidence becomes Needs evidence rather than a guess. Inspect supporting excerpts shows exactly what was matched, and opening it costs nothing. A source match proves a passage exists - never that it is true or complete.

The option: connect a TypeSafe account. In Provider Hub → Quick checks, paste a key and these checks run there instead, typically in a fraction of a second. The key is validated with one real question before it is saved, so a wrong key is refused in a sentence and nothing is stored; Try it runs one sample check and prints the measured time. With an account connected the chip also carries a percentage, and when a check is not sure a second model is asked - if they disagree and you are running with Oversight, Cerevisor asks you, and if nobody is there it takes the cautious branch. The percentage is the model’s own estimate of its own certainty, not measured accuracy, and the app says so where you read it.

What is sent, stated above the key field before you paste anything: Cerevisor will send selected decision inputs, source excerpts, and a short summary of finished runs to TypeSafe to be checked. Your memory is never sent. Connecting is the consent - there is no second opt-in - and the connect, every switch and the disconnect all land in your consent history and your data export. Two switches, both on, turn each purpose off again: Use TypeSafe for decision nodes and edge checks. and Rate how each finished run went with TypeSafe. A run that will reach TypeSafe says so on its pre-run summary. On retention Cerevisor states TypeSafe’s own terms rather than a flattering version: TypeSafe says it does not train on your data. On a standard account it keeps what it is sent. Without a connected account nothing goes anywhere new - a check runs on a model your run already uses, preferring the one that produced the text, so someone on a local model stays entirely on their own machine.

4. Three plans

Plan	What it is
Free	As before, minus the team conversation.
Pro	Everything the paid plan had, except Operator and the team conversation. Up to 5 workflows in one world, and 2 world files.
Founding member	Everything, with no limits.

Everything that already separated free from paid (connected servers, scheduled runs, web search, the Vault, hooks, saving a Power, creating an Organization, the Advanced Loop, smart stop-conditions, the Deep launch preset, Skill Workshop editing, trust-role design, Cursor cloud runs, Loop & Enhance, Endless mode) sits on Pro exactly as it sat on the paid plan.

The Subscription tab now names your plan: Cerevisor Founding member, Cerevisor Pro, Free plan, or the trial countdown. On Pro it adds one line saying what Founding membership adds, with See pricing and Enter a different key beside the buttons that were already there. No price appears anywhere in the app; pricing lives on cerevisor.com.

5. Operator is part of Founding membership

Operator working on its own while you are away is experimental, and it is now part of Founding membership. It is still included in the 7-day trial.

Nothing about the Operator panel closed: the journal, objectives, the review queue and running a tick yourself all stay open on every plan. What stops is the waking up on its own. If it was switched on when your plan changed, it stops at its next wake rather than going quiet without a word, and the panel says why: Operator stopped working on its own because it is part of Founding membership. A run it had already started finishes normally.

6. The team conversation, and what is not the team conversation

The conversation is agents posting notes to each other during a run, asking each other questions, your own free-form notes, pinning a note so it carries into the next run, seeing an earlier run’s conversation on this one, the conversation panel on the canvas, and posting from your phone.

It is not any of these, and none of them changed on any plan:

• Reviews by a second agent. A reviewer still posts its verdict and you still see it, including when the reviewer runs on an outside tool.
• An agent asking you for help, and your answer reaching it. The answer box in the run window’s Board tab still works on every plan, and so does answering from your phone.
• Check by hand, a note’s Why believe this evidence, and the run’s Overview.
• Notes for the agents, the box in the launcher where you tell a run what you want before it starts.
• The notes already on a board. Nothing is hidden. Old runs open and read exactly as they did, and the Board tab still shows everything.

On Free and Pro the Board tab keeps its list and its answer box, and the compose row becomes a one-line card saying which plan the conversation belongs to. Your pins stay saved on the workflow and work again the moment you have Founding membership. An agent running on an outside tool can still ask you for help; that is all its notes file is used for, and its instructions say so instead of describing a conversation it cannot have.

7. The limits never lock your work

A limit stops you creating something new. It never touches what you already have.

A world holding eight workflows opens, edits, runs and exports exactly as before on any plan; only the ninth is refused. Nothing goes read-only, nothing is hidden, nothing is deleted. That was already true of the free limits, and it is true of Pro’s.

Every way of adding a workflow to a world now checks the same thing, which several of them previously skipped: the Add workflow dialog, all three Branch buttons, asking the chat builder to branch one, Duplicate, accepting a workflow Operator drafted for you, and creating or duplicating one from your phone. Each refusal says what your plan allows and what the next one does, and offers a way to upgrade you can ignore.

8. Coming from the trial

The trial is still everything, Operator and the team conversation included.

Two of those things are not part of Pro, so if you buy Pro after using the trial, Cerevisor says so at the moment you enter the key rather than letting you find out when they stop: Your trial included Operator and the team conversation. Those are part of Founding membership, not Pro. It is said once, and you can dismiss it.

Changing plan is the same as it always was: buying a different plan gives you a new key, and entering it in Settings → Subscription replaces the one on this device.

9. Color schemes

The light/dark switch in the title bar is now a Color scheme picker with five choices: Light, Dark, Sand (warm cream and teal), Ocean (deep blue and cyan) and Lavender (pale purple with a violet accent). Sand and Lavender keep light controls and editor styling; Ocean keeps dark ones.

The choice recolors the whole app, the canvas included, and it is remembered: your scheme is applied at startup before anything is drawn, so there is no flash of the wrong colors on launch. A preference that cannot be read falls back to Light rather than to something arbitrary, and an existing saved preference is honored untouched - if you were on Dark, you stay on Dark.

Also fixed

• Visiting Home no longer closes the workflow you had open, and it hides the run window while you are there instead of leaving it floating over the Home screen.
• Right-clicking an agent on the canvas opens its menu without also opening its settings.
• A workflow whose structured output definition was saved in a malformed shape is now repaired on load instead of failing.
• Cut, copy and paste in text fields behave safely everywhere, with a proper right-click menu.
• When switching a connection on or off fails, the app says so plainly instead of showing a toggle that quietly snapped back.

Known limits

• One review setting is quieter without the team conversation. Only when the agent is unsure looks at three signals, and one of them is the agent saying so itself in a note. On Free and Pro that signal is silent and the setting decides on the other two. Always and Off behave identically on every plan.
• A widened conversation range reads as this run only. If a workflow is set to show earlier runs’ notes, that setting stays saved and the run’s own record still states it, but on Free and Pro the run starts with only its own notes.
• The disclosure about the trial can appear once more than it should. A Founding or Pro subscription that has lapsed reads as the free plan, so entering a Pro key afterwards shows the trial sentence above even though there was no trial involved. It is one dismissible message and nothing else follows from it.
• Quick checks: the percentage is the model’s own estimate. Nothing has been measured against ground truth, and the app says so where the number appears. This release makes no claim about how accurate the checks are.
• Quick checks: a standard TypeSafe account keeps what it is sent. Zero retention is an enterprise arrangement with them, not the default.
• Quick checks: a source match proves a passage exists, not that it is true. It establishes where text came from, and nothing more - it does not certify the source’s correctness, originality or completeness.
• Quick checks: the finished-run rating is judged from the run’s own summary - the two lines the run log holds about what worked and what did not, and nothing else.
• Quick checks: loop-exit checks never ask a person. An unsure one ends the loop. A loop’s own stop condition also still does its result check on Anthropic models only, whether or not you connected an account.
• Quick checks: an unsure check’s recorded cost can include a second opinion’s share. When a connected account’s answer is checked against another model’s, both costs ride on the one line - which is why the run’s summary names the account only when every check went to it.
• Quick checks: there is no display on the phone. The companion app shows no chips or totals, and the section in the setup pane only appears once you have at least one AI model connected.
• Quick checks: the pre-run sentence appears whenever this run will reach TypeSafe, including when only the finished-run rating is on.
• Connecting a model: some failure messages still end in technical wording. A failed Codex step can append a raw (Details: …) tail taken straight from the underlying error.
• Connecting a model: the Codex install log has two homes - under Advanced once Codex is found, and inline in the install step before it is, because Advanced does not exist yet at that point.
• Connecting a model, macOS: Node.js installed after Cerevisor launched may still need a Cerevisor restart before Codex can be installed.
• Your first project: the guide opens itself only on a genuinely new installation - no AI model connected, no permission settings chosen, no recent files. An existing installation is never pulled into it; start it from Home instead.
• Everything listed under 3.1.0’s, 3.1.1’s, 3.2.0’s and 3.2.1’s known limits still applies.

What this was verified on

On the finished, combined tree this release is cut from, the whole automated suite, the type checks, the desktop-channel contract check, the mobile checks, the coverage check, a production build and four end-to-end specs were run on one Windows development machine.

On the finished, combined tree: the type checks printed the same three lines as before this release; the desktop-channel contract check passed; 918 of 928 test files and 11,325 tests passed in the full run, and the nine files that did not (eight of the slowest files running out of time under load, and one check that a generated index was current) all passed when re-run on their own; all 310 phone-app tests passed with its type and boundary checks; the production build succeeded; and nine of nine end-to-end checks passed (first launch, the first-project guide, color schemes, and a real recorded run).

Not everything in those runs was green, and none of it is hidden here. In the full suite a handful of the slowest test files ran out of time while thousands of tests were running at once on the same machine; each one passed when it was run on its own, and none of them touches anything this release changed. The end-to-end suite finished with two failures the previous release already has: the two tests that pair a phone and drive a run from it.

Nothing in the sections above has been exercised by hand in an installed app. There was no clean-computer install and no installer of any kind; no real ChatGPT, Claude or Google sign-in was performed; no real TypeSafe key was used; no real license key was activated; and no phone was involved. That covers the whole of connecting an AI model, the first-project guide, quick checks against a live TypeSafe account, activating a Pro key, the upgrade cards, the refusal on a phone and the Operator sentence after a real trial-to-Pro change. Those are written down as unverified rows for a person to run, in docs/plans/2026-09-07-cerevisor-3-1-m2-real-app-checklist.md (K16 onward). Nothing in this release claims otherwise.

Platforms

Available for Windows, macOS (Intel and Apple silicon) and Linux (AppImage and Debian package), all built from the same tagged tree by the release workflow.

Updating

Existing installations receive the update through Cerevisor’s built-in updater, and the installers are available from cerevisor.com/download and the private GitHub release. Your workflows, run history and settings carry forward untouched; no file format changed.