Voice Analyser MCP

MCP server that analyses your published writing and generates executable style guides for voice-matched content creation.

What This Does

Point it at a sitemap, it crawls your articles, runs 16 linguistic analysers, and generates a style guide built from your actual writing patterns. Not generic advice — your phrases, your sentence rhythms, your quirks.

The output is designed for LLM consumption: forbidden word lists, phrase libraries with examples, burstiness targets, and validation checklists that catch AI slop before it ships.

What Changed in v2.0.0

v1.x loaded 5 of the 16 analysis files and produced a 5-part guide heavy on statistics. v2.0.0 loads all 16 and restructures the output around examples.

Before (v1.x): 5-part guide, ~250 lines, stats-heavy, sparse examples After (v2.0.0): 8-part guide, ~500 lines, examples-first with supporting stats

New Guide Sections

| Part | What It Covers | |------|---------------| | 1. Zero Tolerance Rules | Forbidden AI slop vocabulary with alternatives, formal word replacements, punctuation rules | | 2. Phrase Library | 50+ actual phrases from your corpus: opening patterns, equipment references, caveats, transitions | | 3. Sentence Patterns | Length distribution, rhythm variation targets, first-person usage frequency | | 4. Burstiness & Flow | Coefficient of variation, consecutive sentence clusters, rhythm examples from corpus | | 5. Expression Markers | Hollow intensifiers to avoid, hedging/certainty balance, conversational devices | | 6. Argument & Identity | How arguments open/build/close, identity markers (genuine interest, honest obsession, expertise) | | 7. Paragraph Architecture | Opening/closing patterns, transition density, information density scoring | | 8. Stylometric Fingerprint | Function word z-scores, punctuation signature, statistical fingerprint for voice matching |

Each section uses graceful degradation — if the corpus doesn't produce enough data for a section, it's omitted rather than padded with generic advice.

Installation

Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "voice-analysis": {
      "command": "npx",
      "args": ["-y", "@houtini/voice-analyser@latest"]
    }
  }
}

Claude Code

Add to your .claude/settings.json:

{
  "mcpServers": {
    "voice-analysis": {
      "command": "npx",
      "args": ["-y", "@houtini/voice-analyser@latest"]
    }
  }
}

Config file locations (Claude Desktop):

• Windows: %APPDATA%\Claude\claude_desktop_config.json
• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Linux: ~/.config/Claude/claude_desktop_config.json

Restart Claude Desktop after saving.

Requirements: Node.js 18+

Quick Start

1. Collect Writing Corpus

In Claude Desktop or Claude Code:

Collect corpus from https://yoursite.com/post-sitemap.xml
Save as "my-voice" in "C:\writing\voice-models"

The tool crawls your sitemap, extracts clean content from each article, and saves markdown files. Aim for 40+ articles (30,000+ words) for reliable patterns.

2. Run Analysis

Analyse corpus "my-voice" in directory "C:\writing\voice-models"

Runs 16 analysers covering vocabulary tiers, phrase extraction, sentence structure, voice markers, punctuation habits, argument flow, information density, and more. Takes 1-3 minutes depending on corpus size.

3. Generate Style Guide

Generate style guide for "my-voice" in directory "C:\writing\voice-models"

Creates an example-first guide at: C:\writing\voice-models\my-voice\writing_style_my-voice.md

Using the Style Guide

Load the generated guide into Claude conversations:

Load C:\writing\voice-models\my-voice\writing_style_my-voice.md
and use it to write [content type] about [topic]

The guide includes validation checklists. After Claude writes:

Check what you just wrote against the style guide validation checklist.
Report any violations.

What It Catches

• AI slop words (delve, leverage, unlock, seamless, robust) with corpus-specific alternatives
• Em-dashes when your corpus uses hyphens (or vice versa)
• British/American spelling inconsistency
• Generic equipment references ("the product") vs possessive ("my rig")
• Sentence rhythm that's too uniform (low burstiness)
• First-person frequency outside your natural range
• Missing honest caveats and conversational devices

Analysis Output

The tool generates 16 JSON files in corpus-name/analysis/:

| File | What It Captures | |------|-----------------| | vocabulary.json | Word choice, domain terms, British/American markers | | vocabulary-tiers.json | AI slop detection, formality scoring with alternatives | | phrase-library.json | 50+ extracted phrases organised by type | | sentence.json | Length distribution, complexity, burstiness (CV) | | voice.json | First-person usage, hedging, conversational markers | | paragraph.json | Structure, length patterns, transition density | | punctuation.json | Dash types, comma density, parentheticals | | function-words.json | Z-scores for stylometric fingerprinting | | anti-mechanical.json | Naturalness scoring | | expression-markers.json | Hollow intensifiers, hedging/certainty balance | | clustering-patterns.json | Consecutive sentence clusters by length | | argument-flow.json | How arguments open, build, and close | | paragraph-transitions.json | Cross-paragraph connection patterns | | information-density.json | Technical density scoring | | specificity-patterns.json | Possessive vs generic references | | vulnerability-patterns.json | Identity markers (genuine interest, expertise signals) |

Minimum Corpus Requirements

• Minimum: 15,000 words (~20 articles)
• Recommended: 30,000 words (~40 articles)
• Ideal: 50,000+ words

Below 15k words, statistical patterns become unreliable. The phrase library needs volume to find frequently-used patterns.

MCP Tools Reference

collect_corpus

Crawls sitemap and collects clean writing corpus.

| Parameter | Required | Description | |-----------|----------|-------------| | sitemap_url | Yes | XML sitemap URL | | output_name | Yes | Corpus identifier (e.g., "my-voice") | | output_dir | Yes | Storage directory | | max_articles | No | Article limit (default: 100) | | article_pattern | No | Regex URL filter |

analyze_corpus

Runs 16 linguistic analysers on collected corpus.

| Parameter | Required | Description | |-----------|----------|-------------| | corpus_name | Yes | Name from collect_corpus | | corpus_dir | Yes | Directory containing corpus | | analysis_type | No | full, quick, vocabulary, syntax (default: full) |

generate_style_guide

Generates v2.0 executable style guide from analysis data.

| Parameter | Required | Description | |-----------|----------|-------------| | corpus_name | Yes | Name from analyze_corpus | | corpus_dir | Yes | Directory containing analysis |

Development

git clone https://github.com/houtini-ai/mcp-server-voice-analysis.git
cd mcp-server-voice-analysis
npm install
npm run build

Local development mode in Claude Desktop config:

{
  "mcpServers": {
    "voice-analysis": {
      "command": "node",
      "args": ["C:\\path\\to\\mcp-server-voice-analysis\\dist\\index.js"]
    }
  }
}

Known Limitations

• Needs XML sitemap (RSS feeds not supported)
• Works best with single-author content
• Mixed authorship weakens statistical signals
• Heavily edited content produces less distinct voice patterns
• Guide sections are omitted when the corpus doesn't produce enough data for reliable patterns (this is intentional, not a bug)

License: Apache 2.0 Author: Houtini Version: 2.0.0