Scroll MCP Server v0.2

🎯 Vesuvius Challenge submission — Automated quality audit & corpus navigation for the Herculaneum scrolls.

MCP Server for the Vesuvius Challenge papyrological corpus & open data.

Exposes scroll metadata, transcriptions, CT volumes, ML models, and documentation via the Model Context Protocol.

Quick Start

pip install scroll-mcp-server
scroll-mcp-server

Or with Claude Desktop — add to your claude_desktop_config.json:

{
  "mcpServers": {
    "scroll": {
      "command": "scroll-mcp-server"
    }
  }
}

URI Scheme

scroll://                    → Root
├── catalog/scrolls          → Scroll inventory
├── catalog/vocabulary       → Papyrological glossary
├── scroll/{id}/metadata     → Scroll metadata
├── scroll/{id}/transcriptions → Deciphered text
├── scroll/{id}/volumes      → CT scan info
├── model/ink-detection      → ML models
├── docs/tutorials           → Tutorials
├── docs/formats/tifxyz      → Format specs
└── docs/open-problems       → Research challenges

Tools

Tool	Description
`search_corpus`	Search transcriptions & metadata
`get_scroll_catalog`	Filtered scroll inventory
`get_segment_data`	Segment metadata + download links
`get_tutorial`	Tutorial content
`get_format_spec`	Technical format specification
`get_model_info`	ML model metadata
`list_available_models`	Filtered model list

Prompts

Prompt	Purpose
`explain_papyrus_context`	Historical/philosophical context of a passage
`suggest_research_direction`	Research recommendations from corpus state

Architecture

Client → FastMCP → ScrollDataLayer → Catalog (JSON) + Vesuvius (streaming) + Cache (LRU)

Catalogue JSON — Lightweight index (~50 KB) for instant metadata responses
Data Layer — Orchestrator with LRU cache, routes to the right source
Vesuvius Client — Streaming proxy for CT volumes (no local storage)

Development

git clone https://github.com/zengdard/scroll-mcp-server
cd scroll-mcp-server
pip install -e .
pytest tests/ -v

Real Example — Scroll 1 Segment Audit

Un exemple complet avec données réelles du Scroll 1 (PHerc. Paris 4) est dans examples/ :

# Audit du mesh noisy (84k vertices) contre le volume CT
python examples/audit_demo.py --quick

Résultat :

Quality score:    0.431  ← mesh bruité, besoin de raffinement
Mean confidence:  0.942  ← bonne qualité CT
Mean distance:    3.88 voxels
🔴 CRITIQUE — 55650 vertices en drift
🟢 ANCRE — 6418 vertices stables

Voir le rapport complet : examples/AUDIT_REPORT.md

Submission

Ce projet est soumis au Scroll Prize. Voir SUBMISSION.md pour les détails.

Team: Individual — zengdard
Repo 1: https://github.com/zengdard/scroll-mcp-server
Repo 2: https://github.com/zengdard/scroll-label-refiner

License

MIT