Vellum AI Automation Removed: Where Authors Draw the Line
You opened Vellum on Thursday morning to fix a typo in book six, and the app wanted an update. Maybe you'd seen the announcement the night before. Maybe you only noticed when your Threads feed filled with authors swearing they'd never open the app again. By lunchtime the thing they were angry about was gone. You still had a typo to fix.
Here's what happened, and what it says about where working novelists draw the line on AI in their tools. That line is easier to see than either side let on.
What did Vellum actually ship and remove?
On November 4, 2026, Vellum released version 4.2 with a feature called Automation. It let you connect an external AI assistant, such as Claude, to Vellum so the assistant could drive the app through publishing chores. On November 5, Vellum shipped 4.2.1 and removed Automation entirely. The feature lasted about a day.
The details matter, because much of the anger targeted a feature people pictured, not the one that shipped. According to Vellum's own 4.2 announcement, Automation needed a Vellum Press license and a compatible assistant running on your Mac, and Vellum itself had no AI built in. The examples were production work: check that every photo has an accessibility description, regenerate Kindle EPUBs for a folder of books, move an About the Author section into a shared template, reformat text-message scenes with Vellum's Text Conversation feature. Whitney Foster's breakdown of the 24 hours describes it as a Model Context Protocol (MCP) server that did nothing unless you turned it on, and notes that the demo video showed cover swaps and back matter, not a book being written.
Vellum's removal post explained the motive: longtime customers with dozens or hundreds of books, including translations and special editions, had asked for help with repetitive work like updating back matter and building box sets. Then it admitted the miscalculation in one sentence: "we underestimated how strongly many of you would object to Vellum working with AI assistants at all." SMBtech's write-up adds that the email was signed by all three people on the team (Brad, Brad and Katharine) and quotes them saying removal was "the best decision we can make."
Why did an opt-in feature cause this much backlash?
Because for a lot of authors the objection was never about one feature. It was about what the feature signaled. Vellum is the rare piece of author software that people trust almost personally: a small team, a clear price ($249.99 for Vellum Press, $199.99 for Ebooks, per vellum.pub), and a long record of doing one job well. When a tool like that adds an AI hook, the fear isn't "this will write my book." It's "this is step one, and I don't know step two."
The other side was loud too. Dana Sacco's post, Vellum Caved, argues that "I don't want this" turned into "you should not be allowed to have this," and that the people hurt most are authors with big backlists, small presses, and disabled or neurodivergent writers for whom repetitive production work is a real cost. James Blatch called it a tragedy and made the same point more bluntly: critics stopped other authors from using a feature they were never obligated to touch.
Both camps describe something real. Neither quite named the line most authors actually draw.
Where do most authors actually draw the line?
Most authors don't draw it at "AI" versus "no AI." They draw it at two questions: who is holding the pen, and where are the pages going? A tool can pass one test and fail the other, and most of the confusion this week came from treating them as one test.
Call it the pen-and-pages test.
The pen is authorship. Is the model producing the sentences readers pay for, or doing work around those sentences? Swapping a newsletter link in the back matter of fourteen books is not authorship by any definition that matters to a reader. Generating the scene where your detective finally realizes her sister lied about the night of the fire is. Vellum Automation sat firmly on the chores side, and that's why its defenders were baffled.
The pages are custody. When an assistant reads your manuscript to do a task, where does that text go? This is the question the backlash gestured at without quite asking. An MCP connection is just plumbing between apps on your Mac. But whatever the assistant reads goes to whatever model is running the assistant, and if that's a cloud model, your chapters are leaving the machine. That isn't a scandal. It's how cloud assistants work. It does mean "the AI isn't built into the app" and "my book stays on my computer" are two different promises, and Vellum's announcement spoke only to the first.
Run the episode through both questions and the anger makes more sense. Plenty of authors who would happily let a script fix their back matter still flinched at the idea of a cloud assistant reading their whole catalog. With no clean way to say "fine with the pen, not with the pages," it came out as "no AI, ever."
A worked example: the fourteen-book back matter update
Take Marisol, who writes a cozy mystery series about a bookbinder in a lakeside town. She has fourteen books out, each in ebook and paperback. She's just moved her newsletter to a new platform, so every "Join my readers' club" link in every back matter page is now dead.
By hand, each book means opening the Vellum file, editing the link, checking the also-by page, regenerating the EPUB and print PDF, then uploading to KDP and her wide retailers. Call it twenty-five minutes a book if nothing goes wrong. Fourteen books is nearly six hours, and something always goes wrong: book nine still lists the old series order, and book eleven's paperback uses a different trim size she has to re-check.
That is exactly the job Automation was built for. With it, Marisol could have asked an assistant to make the change across all fourteen files and spent her afternoon on book fifteen. It's also a job that touches nothing a reader would ever call writing. On the pen question, it passes easily.
Now the pages question. To update fourteen back matter sections, an assistant has to open fourteen books. If Marisol's assistant runs on a cloud model, some amount of that text goes to the provider. Whether that bothers her depends on her contracts, her feelings about training data, and the retention policy of the assistant she uses. Those are her calls. The point is that she should be the one making them, per task, with the information in front of her.
That's the bit the whole fight skipped. Authors don't want a vendor deciding for them in either direction, by shipping AI they didn't ask for or by yanking a tool they did. They want consent at the feature level, not the vendor level.
What to look for in writing software after Vellum
If you're auditing your own toolchain this week, these questions are worth asking of anything that touches your manuscript, whether it calls itself AI or not:
- Is it off until you turn it on? Opt-in should mean that you see no change, and nothing touches your files, until you deliberately connect something. Vellum's design met this bar, which is part of why its defenders were frustrated.
- Can you choose which model does the work? A tool that only works with one cloud provider makes the custody decision for you. One that lets you pick, including a local model, leaves it with you.
- Can it run with no network at all? This is the only setup where "my manuscript stays on my machine" is literally true for the AI step too. Local models through Ollama or LM Studio are good enough now for a lot of revision and production work, and they cost nothing per call.
- Is the vendor clear about data flow, not just about features? "We didn't build AI in" is a statement about the app. You also want to know where text goes when you do use AI. The NovelCrafter incident earlier this autumn was a reminder that the things a company stores are the things that can leak.
- Does the business model depend on your data? A subscription that funds servers holding your chapters puts the vendor in a different position from a one-time purchase that runs on your desk.
This is where I'll be direct about the product I work on. NovelMage is built around the pages side of that test. It's an offline-first desktop app for Windows, macOS and Linux; manuscripts are stored locally in standard formats, and there are no NovelMage servers that ever handle manuscript content. You can run a local model through Ollama or LM Studio with zero network traffic, or plug in your own Claude, GPT or Gemini key when you decide a cloud model is worth it for a particular job. That choice is yours, per task. The lifetime license is $99.99 once, on up to three devices, so nothing about the business needs your text to live on someone else's machine.
To be fair to Vellum: NovelMage is a drafting and revision workbench, not a book formatter. If what you need is beautiful print interiors and clean EPUBs, Vellum is still excellent at that, and nothing in this week's mess changes it.
What this means for AI-assisted authors right now
First, the workflows haven't gone anywhere. As Foster points out, removing Automation didn't stop authors from using assistants alongside Vellum. It moved that use outside the app, where no company has to own it or defend it. Expect more clicking, not less capability.
Second, your paperwork should catch up with your tools. If you hire help, the question of which AI touches your manuscript belongs in writing. Our post on adding an AI clause to your editor agreement has language you can adapt, and the same logic applies to formatters and virtual assistants.
Third, keep the pen and the pages separate in your own head when disclosure comes up. Using an assistant to fix back matter is a different category from using one to draft chapters, and retailers treat them differently. If you're unsure what that means for your listings, our breakdown of KDP's AI disclosure rules walks through the generated-versus-assisted distinction.
Frequently Asked Questions
Can I still use Vellum Automation?
No. Vellum removed Automation in version 4.2.1, released November 5, 2026, one day after it shipped in 4.2. The rest of 4.2, including the Word Count Graph's new History mode, is not affected. Updating through Check for Updates gets you 4.2.1.
Did Vellum add AI that writes books?
No. By Vellum's own description, AI was never built into the app. Automation was a way to connect an external assistant that you already ran on your Mac, and it required a Vellum Press license. The examples Vellum gave across its two posts were production tasks: accessibility descriptions for images, regenerating EPUBs, shared back matter templates, and box sets.
Does connecting an AI assistant to an app send my manuscript to the AI company?
It depends on the model behind the assistant, not on the connector. A protocol like MCP just lets two programs talk on your machine. But any text the assistant reads gets processed by its model, so if the model runs in the cloud, that text goes to the provider. If the model runs locally, through Ollama or LM Studio for example, it stays on your computer.
What should I use if I want AI help but no cloud at all?
Use a tool that supports local models and run one on your own hardware. In NovelMage you can point the app at a local model through Ollama or LM Studio and work fully offline; the app recommends at least a 7B-parameter model. There's no per-call cost, and nothing leaves the machine.
The line, stated plainly
The week's lesson isn't that authors hate AI, and it isn't that a mob killed a good feature. It's that authors want two promises kept separately: the pen stays in their hand, and the pages stay where they put them. Vellum kept the first promise and said little about the second, and that gap was enough to sink a feature in a day.
If you want an AI workbench that was designed around both promises from the start, download NovelMage and try it free for seven days, with local-model support built in and no credit card required.