Hermes Novel Generator: Multi-Agent Novel Writing as a Skill
AI novel writing tools fall into two buckets. The first bucket is "one big prompt" — describe your idea, hit enter, get a story. The second bucket treats long-form fiction as a process: plan, draft, audit, revise, polish, repeat. The tools in the second bucket are rarer and more interesting.
Hermes Novel Generator (14 stars, MIT) is a skill for Hermes Agent that lives in the second bucket. It splits novel generation across five agent roles and runs the whole thing as a procedural pipeline. You give it a seed concept and an author name. It gives you an EPUB.
The real question is how it stacks up against the two bigger projects in this space: NovelClaw (293 stars, from Harbin Institute of Technology researchers) and Morpheus (25 stars, full-stack with three-layer memory).
The Scene Sandbox Model
The novel generator's architecture is built around a specific claim: characters should actively participate in scene construction, not just get reviewed after the fact.
Here is the per-scene loop:
- Storyteller writes a scene brief — narrative goals, beat structure, which characters are present
- Character Agents each receive the brief plus their own profile. In isolation, they plan their beat-by-beat actions, dialogue, and intentions. They do not see each other's strategies
- Storyteller receives all strategies and weaves them into full scene prose. Conflicts, overlaps, interruptions — the Storyteller resolves them
- Character Agents do a voice-check pass on the woven draft. Each reviews their own lines for authenticity, flags violations
- Storyteller revises only the flagged passages
- Lore Auditor checks continuity, world rules, knowledge boundaries, pacing, themes
- Prose Stylist polishes prose quality, explicitly banned from changing narrative content
- Orchestrator updates world state, character growth, timeline
Each multi-character scene takes about 10 agent calls. Single-character scenes take about 5. A ten-chapter short novel runs 250–400 calls and takes 2–5 hours.
The key insight: characters plan independently, the Storyteller synthesizes, then characters verify. It is a writing workshop, not a factory.
NovelClaw: The Research-Grade Competitor
NovelClaw is the most polished project in this space. Built by researchers at HIT Shenzhen, it ships as a Docker Compose stack with three services: Portal (public entry), MultiAgent (ideation), and NovelClaw (the main writing workspace at port 8012).
What sets it apart:
- A full FastAPI backend with SSE streaming for chapter generation
- A React frontend with chapter workbench, memory browser, knowledge graph, trace replay, quality dashboard
- Dynamic memory architecture (L1/L2/L3/L4) with SQLite + LanceDB vector search
- Human-in-the-loop design — the author keeps steering throughout
- Published as a research artifact with a live demo at colong-idea-studio.cloud
NovelClaw is what you get when an academic team builds a novel writing tool with the seriousness of a software product. It has architecture diagrams. It has deployment guides. It has a GitHub-safe packaging checklist.
The downside: it is a standalone application, not an agent skill. To use it with Hermes, you would need to either run it alongside as a separate service or port its memory-aware workflow into a skill format. It is not a drop-in. It is a platform.
Morpheus: The Three-Layer Memory System
Morpheus takes a different approach. Instead of running agent calls per scene, it builds an elaborate memory architecture and lets the writer work through chapters in bulk.
The memory system has conceptually three layers (L1/L2/L3 active memory) plus runtime state, open threads, and a graph subsystem. When generating a chapter, the MemoryContextService assembles a "generation context pack" containing identity, runtime state, memory compact, previous chapter synopses, and open narrative threads — then hands it to the LLM.
Morpheus also has a chapter workbench, knowledge graph, trace replay, and export pipeline. The tech stack: FastAPI + React 18 + Vite + Zustand. The default LLM path is DeepSeek.
Like NovelClaw, Morpheus is a standalone application. It wants to be a writing studio, not a skill. The memory architecture is genuinely thoughtful, but it also exists inside a product surface that expects you to work inside its UI.
Where Hermes Novel Generator Fits
The novel generator skill is the smallest of the three projects by every measure — stars, code volume, architectural complexity. But it has one structural advantage: it runs inside Hermes.
As a skill, the novel generator turns Hermes into a novel-writing orchestrator. The pipeline is a markdown document (SKILL.md) that the Hermes assistant follows step by step, spawning sub-agents via delegate_task. State lives in novel_state.yaml. Scenes live in markdown files. The dashboard is a simple Python script on port 8420.
This means:
- No separate server to run
- No new tech stack to learn
- Novel state accessible through Hermes's own memory and skill system
- The Scene Sandbox model produces agent calls that are individually inspectable
The trade-off is clear. NovelClaw and Morpheus give you polished UIs and production infrastructure. The novel generator gives you a procedural guide that works within your existing agent. For users who already live in Hermes, the integration path is simpler.
The Numbers That Matter
A fair comparison across the three:
| Aspect | Novel Generator | NovelClaw | Morpheus |
|---|---|---|---|
| Stars | 14 | 293 | 25 |
| Architecture | Skill pipeline | Docker stack | Poetry + Vite |
| Agent model | Scene Sandbox (5 roles) | Multi-agent studio | Agent studio |
| Memory | YAML state file | L1-L4 + LanceDB | L1-L3 + runtime |
| UI | Dashboard HTML | Full React SPA | Full React SPA |
| Integration | Hermes native | Standalone app | Standalone app |
| Output | EPUB | Chapter export | Chapter/book export |
The novel generator does one thing well: it produces a novel from a seed. No more, no less. The other two projects are writing platforms that happen to produce novels.
What Needs Work
The novel generator's time estimates are optimistic. 2–5 hours for a short novel assumes every agent call completes in 30–60 seconds. In practice, rate-limited APIs and context overflow retries can double that.
The dashboard on port 8420 is functional but bare. NovelClaw's workspace and Morpheus's chapter workbench are miles ahead in UX. If you want to review and revise mid-generation, the novel generator expects you to read markdown files.
The "sequential chapter delegation" pattern documented in SKILL.md — one delegate_task per chapter with inline context — is clever for long novels but sacrifices the Scene Sandbox model's character agency. You get speed at the cost of the thing that makes the skill architecturally interesting.
Dependencies are light (pyyaml, ebooklib, Pillow) but the optional cover generation wants diffusers, accelerate, and torch. That is a heavy install for a decorative feature.
Verdict
For someone who wants to generate a novel inside Hermes without leaving their terminal, the novel generator skill is the right tool. Install it with hermes skills install novel-generator and you are writing. The Scene Sandbox model produces prose with genuine character voice because characters plan before the Storyteller overwrites them.
If you want a writing studio with visual chapter management, knowledge graphs, and trace replay, NovelClaw is the answer. It is the only project in this space with research credentials and a deployed demo. Morpheus sits between the two — more polished than the skill, less proven than NovelClaw.
I installed the skill for quick experiments and bookmarked NovelClaw for when the project gets serious. Different tools for different stages of the same problem.
Sources: hermes-novel-generator GitHub, NovelClaw GitHub, Morpheus GitHub, Novel-Claude GitHub