Pirate sources are drying up
As pirate reading sites and file-sharing groups (PDF, ePUB) get taken down en masse, readers are left starved for content. They have money and appetite, but lack a convenient, complete licensed platform to pay for.
NovelKit: AI-powered storytelling — a studio-grade engine for long-form fiction
Build a licensed catalog with real-author quality, reviewed to studio standard.
No more soulless, logic-drifting AI prose. NovelKit runs a Multi-Agent architecture kept in perfect sync with each project's Story Bible and World Rulebook. Every chapter of a long-running series passes a strict quality gate (85/70), stripping out AI-flavor to deliver natural, deep, genuinely engaging writing on par with real authors. Scale your line-up with precision and total continuity.
- 200+
- Chapters per week
- 85/70
- Quality gate thresholds
- 6
- Dedicated AI specialists
- 300+
- Chapters per series
agent-overview.live
Review · QA runningControl PlaneReview
- CACharacter ArchitectDoneIdle1/1idle
- WBWorld BuilderDoneIdle1/1idle
- QAQuality AuditorRunningch.0015.review14/172 queued
- LKLãng KháchNextch.0015.sync14/173 queued
- PWPlot WeaverQueuedch.0016.outline16/182 queued
- PRWProse WriterQueuedch.0016.write15/172 queued
Current focusLive
Quality Auditor is processing chapter.0015.reviewVietnam's licensed fiction market
Demand for licensed fiction is rising fast, but clean, high-quality supply is scarce.
Shifting reader taste
Vietnamese readers, especially Gen Z and Alpha, are tiring of repetitive foreign tropes. They crave works with Vietnamese settings and voice, drawing on local history, mystery, or homegrown fantasy.
Chat AI struggles with continuity in long-form
AI is strong at idea prompts and short passages, but in long-form it tends to lose the thread, forget world logic (canon), and repeat itself (AI-flavor).
What makes NovelKit different
Controlled by a deterministic Control Plane and a Multi-Agent architecture.
NovelKit doesn't make you 'chat' with an AI to scrape together paragraphs — it runs an automated production line with strict review between dedicated AI specialists (Plot Weaver, Prose Writer, Quality Auditor).
- A professional production line
- Quality gate before publishing
- An auditable handoff package
- Scale across the whole line-up
Multi-agent orchestration
Six AI specialists, one control plane, zero canon drift.
Every request passes through a single door — the delegate_tool seam. The orchestrator reads canon, assigns the right specialist, then blocks at the quality gate before writing. No specialist calls another on its own, so you can always trace who did what, and when.
Delegation flowhub-and-spoke
- RequestCLI · API · cron
Every chapter command enters one shared queue.
- OrchestrateOrchestrator · Lãng Kháchdelegate_tool
Reads canon, picks the specialist, keeps handoff order.
- Execute6 genre specialists
Character · World · Plot · Prose · Quality · Sync.
CACharacter ArchitectCharacter profiles & relationships
WBWorld BuilderWorld & genre rules
PWPlot WeaverOutline along the arc map
PRWProse WriterWrites chapters + recall
QAQuality AuditorScores 85/70, blocks errors
LKLãng KháchSync & gate keeping
- Quality gatereview 85 / 70 → sync
Only a passing score is written to canon; a fail returns for a rewrite.
Production Pipeline
A deterministic DAG is the single source of order — two modes on the same skeleton.
Legacy always runs even with every long-form flag off, so rollback is safe. Compass adds two scroll tiers (Volume/Act) to run long-form novels without replacing the DAG with a free LLM loop.
Legacy · full_plan / rolling
- 01
outline
Plot Weaver expands the chapter outline.
- 02
write
Prose Writer writes the draft.
- 03
self_check
Pure function blocks AI-flavor + repeated sentences.
- 04
review
Quality Auditor scores 85/70.
review pass → sync · soft/hard fail → back to write
Compass · long-form
- 01
bootstrap.compass
Builds compass.md + a multi-Act arc-map skeleton grouped by Volume.
- 02
expand Act
advance_expansion sets the next Act [start..end], capped at target_chapters.
- 03
chapter loop
outline → write (+recall/style) → self_check → review → sync.
- 04
Act / Volume boundary
arc_summary → Act done → expand next; at Volume end, volume_summary + update_compass.
P13 gating: chapter.N is ready only when N ≤ expanded_through_chapter. Legacy: gating always true.
Technical System
NovelKit solves long-form with architecture, not just prompts.
Core philosophy: "keep the writing brain, replace the technical skeleton." Literary knowledge is packaged into Hermes Custom Tools, plugins and skills; runtime, provider and control plane run on Hermes infrastructure.
L1Surface
CLI, FastAPI gateway, cron and React Studio. A thin layer that decides nothing — everything goes through delegate_tool.
L2Orchestration · hub-and-spoke
Hermes agent loop → delegate_tool (1 seam) → Orchestrator Lãng Khách → genre squad. Reads canon before assigning work.
L3Control plane · deterministic
PipelineEngine (DAG, breaker, gating, expansion) + PipelineStateStore (atomic, digest, checkpoint) + AutoNovel adapter.
L4Capabilities · Tools + Plugins
14 custom tools (novelkit_*), pure/idempotent, returning only facts, plus 2 single-select plugins: context_engine and memory.
L5Canon store · file-first
PROJECT_DNA, database, outlines, chapters, reviews, summaries, memory, style_vault, logs, .commits — canon is the source of truth.
File-first canon
Canon is stored as files; the context-engine ranks canon above derivatives even at lower relevance (P5).
8-category memory
8 itemized categories in per-novel SQLite, rotation at 3500 words, isolated per story (P7/P8).
Durable control plane
Atomic writes (temp+fsync+rename), digest + optimistic version, transaction manifest for sync recovery.
Quality gate 85 / 70
REVIEW_PASS_SCORE=85, SOFT_FAIL=70, verdict override; self_check blocks AI-flavor + repeats before review.
Memory Stack
Five memory layers A–E, eight categories, isolated per story.
Memory is stored itemized in a separate SQLite per novel (memory isolation, P7) and rotates at a 3500-word threshold (P8). The memory plugin is a single-select plugin alongside context_engine. This is structured memory, not a true graph database — relationships are inferred by tools at query time.
Layer A · Canon5 layers A–ELayer E · Curated
SQLite per-novel · isolation (P7) · rotation 3500 words (P8).
8 itemized categories
character_stateCharacter state per chapter
story_factsLocked-in story facts
world_rulesWorld rules, genre constraints
timelineTimeline, event order
open_loopsOpen threads awaiting payoff
reader_promisesPromises made to readers
relationshipsRelationships between characters
minor_castNamed minor-cast ledger, deduped
Correctness Properties
24 correctness properties P1–P24, verified by property-based testing.
Each property has a test in tests/ run with Hypothesis. They cover DAG order, canon authority, idempotency/migration, long-form continuity and adaptive operation.
DAG order & operation
P1 ordering · P3 breaker · P4 rolling window · P12 resume · P13 gating
Canon & authority
P5 canon > derivative · P6 language guard · P7 memory isolation
Idempotent & migration
P2 gate monotonic · P8 rotation · P11 idempotent sync · P24 lossless migration
Long-form continuity
P15 summary coverage · P16 arc-map · P17 recall · P18 minor-cast · P19 style stats
Service Lines
Six genre lines, each with its own rulebook.
Pick a genre for your catalog needs. NovelKit locks in the matching world rules, character state, plot threads, style profile and quality checklist for each line.
Business Connect
Start with a sample, not a contract.
Send us the genre, target output and licensing model you need. We'll build a sample with its story bible and pipeline logs so you can assess quality first — expand the catalog only when you're confident.
Answer Engine Notes
Quick answers for partners looking for a licensed-fiction studio.
It's a long-form fiction production system built on a Multi-Agent architecture paired with a deterministic technical backbone (Control Plane). Rather than a single chat tool, NovelKit is an automated line from outlining to detailed writing to quality review for mass publishing.