HNPX SDK
Enables AI-guided hierarchical fiction writing through structured XML documents. Supports step-by-step narrative expansion from book-level planning down to individual paragraphs, with custom Roo Code modes for collaborative story development.
README
HNPX SDK
A collection of tools for HNPX - a hierarchical XML-based format for planning and writing fiction. This project introduces the HNPX format and provides tools for creating, editing, and managing HNPX documents.
Origin
LLMs can't write a good story. Well, technically they can write a good story: they'll generate thousands of words that follow narrative conventions. But they can't write your story, the specific story only you have in mind. That gap between "AI can write fiction" and "AI can't write the fiction I want" is the core of the workflow described below.
Ask an LLM to write a whole chapter, and it'll speedrun through the plot in about 300 words. Try going paragraph by paragraph, and you'll find the AI doesn't know when to stop one scene and start the next.
This led to experimenting with hierarchy: first generating a detailed chapter plan with sections, then asking the LLM to write each section individually based on that plan. This approach improved structure and word count: the chapter speedrun problem was less severe, and scene boundaries were clearer. This hierarchical method required automation, which I implemented through simple Python scripts. However, a fully automated process without guidance led the LLM to write nonsense, not at all what I wanted. The text became filled with errors, and the problem was even more pronounced when writing in Russian, where general LLMs produce noticeably poorer text quality.
The solution evolved into HNPX, a narrative planning format that structures the writing process from book-level overview down to atomic paragraphs. It allows for step-by-step expansion, maintaining context and control at each level while keeping human guidance central to the process.
To bring HNPX to life, I considered building a custom UI or VS Code extension, but that would have been too difficult: creating a whole AI agent system from scratch. Instead, I realized Roo Code already provided the perfect foundation. Its flexible mode system meant I could build an MCP server and create custom modes to use it, creating exactly the workflow HNPX required. HNPX structures the writing process so that LLMs write under human guidance at each step, transforming them from autonomous storytellers into responsive collaborators.
TL;DR: HNPX defines a structured writing workflow where LLMs write under human direction at each level, turning AI from an autonomous storyteller into a guided collaborator.
What is HNPX?
HNPX (Hierarchical Narrative Planning XML) is a structured XML format designed for planning and writing fiction. It organizes narrative content into a strict hierarchy from book-level overview down to individual paragraphs, enabling a step-by-step expansion approach that maintains context and control throughout the writing process. See docs/HNPX.md for a full specification.
The format supports both high-level planning (summaries without full text) and complete documentation (full prose ready for rendering), making it flexible for writers at any stage of their creative process.
See this file as an example of an HNPX document.
Features
- MCP Server: Model Context Protocol server for AI agent integration
- CLI Tools: Command-line interface for some tasks
- Roo Code Integration: Custom modes for collaborative fiction writing
- HNPX Editor Mode: Interactive mode for creating and editing hierarchical narrative documents
- Bible Writer Mode: Specialized mode for creating fiction reference documents
MCP Tools
The HNPX SDK provides a comprehensive set of MCP (Model Context Protocol) tools that enable AI agents to fully manipulate HNPX documents, from creating new documents to editing individual paragraphs and restructuring the narrative hierarchy. See docs/mcp-tools.md for complete documentation.
Installation
Install the HNPX SDK:
pip install -e .
# or via uv
uv pip install -e .
Roo Code Modes Setup
To use the custom Roo Code modes for HNPX editing, place .roomodes into your project root:
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。