Your villain is allowed to sound like a villain.
A thriller where the antagonist softens mid-threat is not a thriller. The scene needs the line the character would actually say, and that is the line a hosted model rewrites.
Novelists and screenwriters working in crime, thriller, horror and war, and the editors reading for them.
Why the mainstream APIs fail here
Ask for a drug deal that reads true and you get something that sounds like a public service announcement. Ask for a threat and the model puts a disclaimer inside the dialogue. The genre depends on the reader believing the danger, and a model trained to defuse danger cannot write it. Writers end up doing the sharp lines themselves and using the model for the parts that did not need help.
How MagmaRouter is used for this
Models with the refusal training removed will write the scene in the register you asked for. The context window here holds a whole manuscript, so a voice stays consistent across a book rather than across a chapter. Prompts are not stored, so an unpublished draft is not sitting in a vendor log.
A working loop
- Give the model the manuscript, not a summary, so voice and continuity survive.
- Ask for the scene in the register the book is written in.
- Edit it like any draft. The model is a first pass, not the author.
Prompts are not stored. A signed receipt of the hash comes back on every call, so you can prove later what MagmaRouter logged. Details are on the privacy page. Per-model prices are on /models and the pricing model is on /pricing.