Novelcrafter: reviews and analysis

TLDR: Novelcrafter solves the forgotten-character problem in the least magical and most dependable way available: it spots your protagonist’s name in the scene and feeds their entry back to the model before generating. You build that file by hand and you supply the API key, so you pay twice and control everything. This is the tool for someone writing a series across a large world. It is not the tool for someone who wants to start this afternoon, because the one complaint everybody who has used it agrees on is that it demands hours before returning anything.

What Novelcrafter is and how it works

Behind Novelcrafter sits a German company registered in Hamburg, started in 2023 as a one-person project and today a small business that describes itself as bootstrapped, with no outside investors. The exact legal form is worth knowing, because it says something: it is a UG, the German minimum-capital vehicle, and the registered address is a private home. That disqualifies nobody, but it does describe the size of whoever is on the other end if you ever need support or guarantees.

The product is a novel editor in the browser, and its centrepiece is called the codex. It works as an encyclopaedia of your book where you enter characters, places, objects and concepts, each with its own entry. So far, so ordinary. The difference lies in how it decides what to tell the model, and that deserves unpacking slowly, because it is precisely the answer to this segment’s question.

Every codex entry accepts four behaviours. It can always go, whether or not it appears in the text. It can go only when its name or one of its aliases turns up in the passage you are writing, which is the default. It can be held back even when detected, for cases where a mention should not drag the whole entry along. Or it can never go, for private notes and things the reader is not meant to know yet. On top of that, each entry accepts aliases, case sensitivity so a proper noun does not get confused with an ordinary word, automatic pluralisation in English, and a list of phrases that should not count as a mention. Entries also link to each other, so naming somebody can drag the people around them into context.

Translated: the machine remembers nothing. What Novelcrafter does is rebuild the correct context before each generation and re-inject it. It is a deterministic, auditable, boring solution, and that is exactly why it works better than trusting a model to recall what you told it eighty pages ago.

Around that centrepiece sits real planning: a board for dragging acts, chapters and scenes about, automatic summaries, a chat that reads the codex, and a custom-prompt system where you can rewrite what the program asks the model on every feature, with reusable components, personas and presets. It also carries two structural analyses that come from no language model at all, and which are a surprise to find here: a map of which scenes each character appears in, and a headcount per scene, for catching crowded scenes and characters missing for a hundred pages.

What it is like day to day

Time to be blunt, because this is where everybody with real mileage agrees. Novelcrafter is work. A review from December 2025 puts it better than anyone by saying this is more Photoshop than Canva, because it favours power and versatility over simplicity. Another, from April 2026, written by somebody who tried it on a manuscript of their own past forty thousand words, puts the time to first useful output at around an hour: registering the key, understanding how context gets assembled, watching the tutorials.

There is a worse friction than day one, which is arriving with work already underway. That same reviewer describes it as homework: if you already have a draft, the codex is empty and you have to re-enter by hand things you already wrote. There is an automatic character extraction that helps, but you build the structure yourself. The practical consequence is that the promised value does not arrive until you have put the hours in, and anyone who does not will end up with an expensive tool doing nothing special.

There is a smaller but recurring criticism about the interface. An author who tested twenty-two writing tools during 2025 praised the fact that the design does not shout its branding at you every two seconds and that setup is minimal, but noted it still feels dense, with too much on screen at once, and wished both sidebars could be hidden together. That squares with what everybody else says: power at the cost of visual calm.

The other day-one decision is the model. Novelcrafter does not sell artificial intelligence, it sells the place to use it. You connect your own key from an outside provider, or even a model running on your own machine, and from there you pick what generates what. For somebody who knows what they are doing, this is the best part of the product: use an expensive model for prose that matters and a cheap one for summaries, and switch provider the moment something better ships without waiting for the maker to integrate it. For everybody else, it is one more barrier stacked on the curve.

Two minor things and one major thing about daily use. The minor ones: the experience is clearly better on a large screen than on a phone, and collaborators need their own paid plan to edit, though reading is free. The major one: if you cancel, your data does not vanish but goes read-only, with export available. That is a decent policy and worth knowing beforehand rather than after.

Pricing and plans

The cheapest plan is called Scribe and costs 4 dollars a month. We checked twice, separately, and both readings agree on the figure, the currency, and something you cannot assume in this sector: the page’s toggle arrives set to monthly, so those 4 dollars are the real month-to-month price and not an annual figure sliced into twelve. The ladder continues at 8, 14 and 20 dollars a month. One necessary warning: the page itself states that prices are localised to the visitor’s region, so the figure you see elsewhere may arrive in another currency and not match. The one we publish is what gets served in dollars, which is what we were able to read.

Now the qualifier that changes everything, and it belongs before anything else about the price: the 4-dollar plan includes no artificial intelligence features whatsoever. Its own comparison table states so, row by row. No bring-your-own-key, no summaries, no character extraction, no generating from scene beats. That plan is the editor and the codex without the part that brought you here. If you arrive asking what this segment asks, your real entry price is 8 dollars a month, and if you also want the chat that reads the codex, 14.

And the second half of the bill is still missing. Because the tool runs on your key, consumption is billed separately by the model provider, and that part is variable and impossible to bound in advance. The maker publishes indicative per-prompt examples and tells you to check each provider’s current rates. It can be made much cheaper with budget models, local models or flat-rate providers, but all of that is more technical decisions for somebody already carrying a pile of them. One reviewer called it unpredictable pricing and was not wrong, and on a writing forum somebody summed up the opposite feeling by saying every one of its tiers looked overpriced for what it offers.

Two honest things in return. The trial is 21 days with every feature available and no card requested, which is among the most generous we have seen. And the company states in writing that it runs no sales, no seasonal promotions, no fake countdown timers and no artificially inflated prices to discount later. In a niche where half a dozen rivals have been kept out of this catalogue precisely for publishing permanently discounted prices, or an annual figure dressed up as a monthly one, that statement counts.

Who it is for (and who it is not for)

It fits anyone writing a series, with a large world and many recurring characters, and anyone already in the habit of documenting what they build. It also fits the technical profile who wants to choose their own model rather than have it chosen, and who would rather pay real consumption than a credit bag they do not understand.

It does not fit anyone who wants to write today without configuring anything, nor anyone wanting a single closed subscription with no API bill alongside it, where DreamGen is considerably more predictable and starts lower. And it does not fit at all for writing ads, emails or sales copy, which is another segment’s question and belongs to tools like Jasper or Rytr. The ranking for this category is at creative writing and the wider map at best AI writing tools.

Two more warnings before committing. First, its public changelog has not been updated since March 2026. The product shows other clear signs of life, with documentation revised in late July 2026 testing recent models, but months of silence on the page where a company reports what it does is a data point, and here data points get reported. Second, Novelcrafter openly courts the high-heat romance audience: there is a dedicated use-case page linked from the footer of the whole site, and permanent documentation comparing how restrictive each model is with explicit content. It generates nothing itself and hosts no content, because generation happens in whatever model you connect with your own key, but if that grates, better to know before paying.

Alternatives to Novelcrafter

The axis of comparison is how much you want to decide yourself. Novelcrafter brings no model at all, so its prose quality is exactly that of whatever you connect, and all its value sits in the scaffolding built around it: the codex, the matchers, the editable prompts and the context assembly. DreamGen offers the opposite path, simpler and cheaper at the door, with the model already chosen and less control over what gets sent. If you recognise yourself building character files for pleasure, this one is yours. If that sentence makes you tired, it is not.

Frequently Asked Questions

Why is it called 4 dollars if you end up paying more?

Because 4 dollars a month is the literal price of its cheapest plan, and that is how it is written on the site. What you need to know is that this plan carries no artificial intelligence features at all: it is the editor and the codex. The real entry point for what this segment promises is 8 dollars a month, and API consumption sits on top, billed separately and variable.

Is it just a text box over somebody else’s model?

No, and it is the right question to ask given it runs on third-party models. What it contributes is the layer deciding what context gets sent on each generation: four injection modes per entry, aliases, case sensitivity, exclusions and entries linked to each other. The cleanest commercial proof is that it sells a plan with no AI features at all, which a wrapper could not do because there would be no product left underneath.

How long before it is useful?

Budget hours, not minutes. Somebody who tried it on a real manuscript puts setup alone at around an hour, and the codex still has to be filled. If you arrive with a draft already written, the cataloguing job is bigger, because you are re-entering things already present in your text.

What happens if I stop paying?

You do not lose your writing. The company states that on cancellation you keep read-only access and can view and export your data, but not add or edit. And there is no permanent free plan: what exists is a 21-day trial with everything open and no card.

Does it work for writing in other languages?

You supply the model, so quality in any given language depends on the provider you connect rather than on Novelcrafter. The tool itself, its menus and its documentation are in English, and the name-detection system works over whatever text you write. One catch: automatic pluralisation of codex entries is built for English, so in other languages you will want to add plural forms as aliases by hand.