Nium-Wiki

English | 中文

A skill for AI coding tools (e.g., Claude Code) that turns codebases into high-quality wikis. Auto-analyzes structure, generates docs with diagrams and cross-references. Inspired by DeepWiki, ZRead, Code Wiki.

Features

• 🚀 Deep Code Analysis: Semantic understanding of your codebase, beyond syntax parsing
• 📊 Mermaid Diagrams: Auto-generated architecture, data flow, and dependency diagrams
• 🔗 Cross-linked Documentation: Bidirectional links between docs with source traceability
• ⚡ Idempotent Incremental Updates: Safe to re-run anytime — SHA256 diff + dep-graph propagation rewrite only the pages affected by code changes, preserving every other doc untouched
• ➕ Additive by Design: New modules, new secondary languages, and hand-edited domain pages coexist across re-runs — no full-wiki rebuild, no clobbering existing content
• 💡 Compiled, Not Retrieved: Aligned with the Karpathy LLM Wiki pattern — docs are compiled upfront, updated incrementally
• 🌐 Multi-language Support: Works with JS/TS/Python/Go/Rust/Java and 10+ languages
• 🧭 Import-Graph Module Discovery: Language-agnostic module discovery via dependency graph, paired with two-phase fact extraction for accurate API docs
• 🔒 Fully Offline: Zero external dependencies, works in air-gapped environments
• 📝 Professional Output: Enterprise-grade documentation with automated quality auditing

Quick Start

Installation

Before using Nium-Wiki, you need to add it as a skill to your AI coding tool:

npx skills add https://github.com/niuma996/nium-wiki --skill nium-wiki

Nium-Wiki works as an AI coding tool skill (e.g. Claude Code). Just tell the AI:

# In Claude Code, say:
> generate wiki
> create docs
> update wiki
> rebuild wiki

The skill will automatically run the full workflow: init → analyze → deep code reading → generate docs → build index → audit quality.

For incremental updates after code changes:

> update wiki

Changed files are detected via SHA256 hashing. The incremental command combines diff analysis, dependency graph traversal, and doc-to-doc propagation to pinpoint affected wiki pages — regenerating only what changed, nothing more.

Generated Output Structure

.nium-wiki/
├── config.json              # Language & exclude settings
├── meta.json                # Version, timestamps, stats
├── cache/
│   ├── structure.json       # Project structure snapshot
│   ├── source-index.json    # SHA256 file hashes (change detection)
│   ├── doc-index.json       # Source ↔ Doc bidirectional mapping
│   ├── dep-graph.json       # Import/require dependency graph
│   └── facts/               # Per-module fact extraction cache
└── wiki/                    # Generated documentation
    ├── index.md             # Project homepage
    ├── architecture.md      # System architecture + Mermaid diagrams
    ├── getting-started.md   # Quick start guide
    ├── doc-map.md           # Documentation relationship map
    ├── api/                 # API reference docs
    ├── <domain>/            # Domain-organized module docs
    │   ├── _index.md        # Domain overview
    │   └── <module>.md      # Module documentation
    └── ...

Multi-language Support

# Initialize with primary (first) + secondary language
npx nium-wiki init --lang zh/en

Secondary language docs are generated in wiki_{lang}/ directories (e.g. .nium-wiki/wiki_en/), mirroring the same structure as wiki/.

Local Preview

# Install
npm install -g nium-wiki

# Start local documentation server (default port 4000)
npx nium-wiki serve

# Specify port
npx nium-wiki serve --port 3000

# Specify wiki directory
npx nium-wiki serve .nium-wiki/wiki

Open http://localhost:4000 in your browser to preview the generated docs, with full-text search, sidebar navigation, Mermaid diagram rendering, and a source code drawer that opens when you click any source file link in the wiki.

Relationship Graph

Click the Graph button in the top-right corner of the preview server to open an interactive relationship graph at /_graph.

The graph visualizes all nodes and edges extracted from your project:

• Blue nodes — source code files (.ts, .js, .py, etc.)
• Orange nodes — documentation files (.md)
• Edges — three relationship types:
  • import — one source file imports another
  • refers — a doc file references a source file (or vice versa)
  • links — a doc file links to another doc file

For large projects (> 500 nodes), the graph loads docs and their directly linked source nodes first, then expands outward via import edges until ~500 nodes are visible. Click any source node to reveal its hidden import neighbors on demand.

Click a node to highlight it: connected nodes stay at full color, unrelated nodes dim. The detail panel on the right shows incoming and outgoing edges for the selected node.

Wiki Generation Showcase

See claude-code-sourcemap-wiki for a live example of a full project generated with Nium-Wiki.

Configuration

After initialization, a default config is generated at .nium-wiki/config.json:

{
  "language": "zh",
  "exclude": [
    "node_modules", ".git", "dist", "build",
    "coverage", "__pycache__", "venv", ".venv"
  ],
  "useGitignore": true
}

Field	Type	Default	Description
language	string	en	Documentation language. Supports zh, en, ja, ko, fr, de. Use / for multi-language, e.g. zh/en (primary Chinese, secondary English)
exclude	string[]	see above	Directories to exclude from code analysis and doc generation
useGitignore	boolean	true	Automatically read .gitignore directory rules and merge into the exclude list

In addition to custom excludes in config.json, the tool has built-in common exclude directories (.git, .idea, .vscode, node_modules, dist, etc.) and language-specific rules from each handler (e.g. Python's __pycache__, Go's vendor), so no manual configuration is needed for those.

Offline-First Design

Nium-Wiki is designed to work completely offline with zero external dependencies, perfect for enterprise internal networks and air-gapped environments:

Preview Server

• ✅ All frontend assets (Docsify, Prism.js, Mermaid) are bundled locally
• ✅ No CDN requests or external API calls

Token Cost

Nium-Wiki is designed to leverage the AI coding tool's existing understanding of your project (from the Explore process), guiding it to output documentation in a structured way — rather than analyzing the entire codebase from scratch.

Scenario	Token Cost	Notes
First generation (with Explore context)	Moderate	The AI tool already understands the project structure, jumping straight to doc generation. Cost depends on project size
First generation (fresh project)	Higher	Requires a full code reading and analysis pass — essentially a deep Explore + doc generation
Incremental updates	Very low	Detects code changes via SHA256 hashing, pinpoints affected docs for targeted updates
Multi-language translation	Very low	Translates after primary language docs are complete, also supports incremental update logic

For projects where you've been using an AI coding tool for a while, the first generation cost is typically reasonable. Day-to-day usage is dominated by incremental updates, which cost very little.

Upcoming Features

• Continuously optimizing documentation quality while reducing model interaction rounds
• Centralized documentation management service
• Intelligent search tooling (cross-project)
• Karpathy LLM Wiki Schema Layer: extend the existing module-facts schema to cover wiki organization rules (domain partitioning, section contracts, cross-doc relations) and drive section-level incremental updates from it

Use Cases

• Enterprise Documentation: Generate comprehensive docs for internal projects
• Open Source Projects: Maintain up-to-date documentation automatically
• Code Reviews: Visualize architecture and dependencies
• Onboarding: Help new developers understand the codebase
• Air-gapped Environments: Works completely offline

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

License

MIT

Links

• GitHub Repository
• Issue Tracker

---

Generated with ❤️ by Nium-Wiki