A local knowledge base and AI Agent context (MCP) service for Project SEKAI fandom.
English | 中文
SekaiSync is built specifically for Large Language Models (LLMs) and coding Agents (such as Claude, Cursor, and ChatGPT). It syncs official Master Data, multi-region localized terminology, and community story texts for Project SEKAI (PJ:SEKAI) into a local store/. Exposing retrieval capabilities through MCP (Model Context Protocol) or CLI commands, it ensures AI Agents rely on ground-truth local data when answering game-related questions—eliminating AI hallucinations at the source.
- ⚡ Zero External Dependencies: Built entirely using Python 3.10+ standard library. Lightweight, pure, and ready out of the box.
- 🌐 5-Region Data Alignment: Full support for Master Data synchronization and cross-lingual translation mappings across JP, EN, CN, TC (Traditional Chinese), and KR servers.
- 🤖 MCP-Native Agent Readiness: Built-in MCP stdio and Streamable HTTP services for seamless integration with Claude Desktop, Cursor, and custom AI workflows.
- 🛡️ Strict Data Boundaries: Adheres to a "report unknown when uncovered" principle to prevent AI Agents from hallucinating nonexistent lore.
- 📰 Flexible Data Extensions: Supports official news announcements sync and TOS-compliant community story text crawling.
git clone <repo-url> sekaisync
cd sekaisync
pip install .💡 Since there are zero third-party PyPI dependencies, you can also run commands directly via
python -m sekaisync <command>without installation.
# Initialize local store directory (v2)
python -m sekaisync init
# Sync Master Data across all 5 regions
python -m sekaisync sync --regions jp,en,cn,tc,kr
# (Optional) Sync official news & announcements
python -m sekaisync news sync# Cross-lingual name resolution (e.g., resolve "星乃一歌" to English)
python -m sekaisync resolve --query "星乃一歌" --target-language en
# Targeted lookup with specific language
python -m sekaisync lookup --query "Hoshino Ichika" --language zh_tw
# General knowledge base query
python -m sekaisync query --query "Hoshino Ichika"
# Check knowledge base and sync status
python -m sekaisync status
python -m sekaisync kb-statusSekaiSync provides a standard Model Context Protocol (MCP) implementation to easily plug into LLM workflows:
Start the MCP service via standard input/output (stdio):
python -m sekaisync serve-mcpStart the HTTP + MCP Streamable service:
python -m sekaisync serve-http --host 127.0.0.1 --port 8787For detailed Agent configuration examples (e.g.,
claude_desktop_config.json), please refer toagents/README.md.
To ensure accurate output from AI Agents, SekaiSync strictly defines its data storage boundaries:
| Category | Coverage | Description |
|---|---|---|
| 5-Region Master Data | ✅ Included | Official metadata for cards, events, songs, characters, etc. |
| Localized Terminology | ✅ Included | Cross-lingual mappings for characters, songs, and terms across JP/EN/CN/TC/KR. |
| Official Announcements | ✅ Included | Synced via news sync. |
| Community Story Text | Text only. Requires custom endpoints and running crawl. |
|
| Multimedia Assets / Runtime Data | ❌ Excluded | Does NOT store images, audio, Live2D assets, chart files, or real-time player data. |
📌 Anti-Hallucination Design: When queried information is outside the current
store/coverage, SekaiSync explicitly returnsnot covered, instructing the Agent to honestly respond that the information is unknown rather than inventing facts.
SekaiSync includes an experimental terminology pipeline that cuts proper nouns and content words directly out of story text (JP/CN/EN/TC/KR), then aligns them across the five languages:
- Zero-dependency tokenizer: language-block splitting + bidirectional maximum matching (Bi-MM) with unsupervised word discovery (frequency × PMI cohesion × boundary entropy), inspired by CJK segmentation research (arXiv:2407.19400, 1905.01964).
- zh-first extraction: simplified-Chinese-priority pipeline (
sekaisync terms zhfirst) that filters the 26 playable protagonists, speaker labels and function words, then inherits official five-language names when a term hits the glossary. - Cross-language penetration: same-position alignment (
sekaisync terms penetrate) and released/unreleased tag clouds (sekaisync tag-clouds). - Honest gating: statistical alignment is only adopted with ≥2-story evidence (containment ≥ 0.30); low-confidence terms stay language-local instead of producing wrong translations.
⚠️ Status: experimental. Recall on a 487-word manual gold set is ~81%, and fragment filtering still relies on LLM judgement (--llm-config) for 4-character colloquial fragments that rule-based filters miss. Do not treat extracted terms as an authoritative lexicon yet — verify against the official glossary when accuracy matters.
The sync command (Master Data) fetches from public community repositories by default and requires no extra setup.
To use crawl (community story text crawling) or custom news sources, configure your endpoint URLs in settings.json:
{
"version": 2,
"sites": [
{
"id": "altsource_sv",
"backend": "sekai_viewer",
"enabled": true,
"master_base": "<your_master_endpoint>",
"asset_base": "<your_asset_endpoint>",
"asset_buckets": { "jp": "<your_jp_asset_bucket>" },
"i18n_base": "<your_i18n_endpoint>"
},
{
"id": "altsource_ms",
"backend": "moesekai",
"enabled": true,
"site_base": "<your_site_address>",
"sitemap_url": "<your_sitemap_url>",
"metadata_bases": ["<your_metadata_endpoint>"],
"asset_bases": ["<your_asset_endpoint>"],
"translation_base": "<your_translation_endpoint>",
"news_base": "<your_news_endpoint>"
}
]
}Compliance & Terms Notice:
- Please ensure you read and agree to the game's Terms of Service, and respect target sites'
robots.txtand content signals. SekaiSync crawls plain text only.- Executing the
crawlcommand (via the--accept-tosflag or interactive prompt) constitutes your explicit agreement to the game's Terms of Service.- Compatibility with the two mainstream public WebDB systems for this game does not imply that SekaiSync endorses or suggests connecting to any particular instance.
Run unit tests:
python -m unittest discover -s tests -t .Build Wheel package:
python -m pip wheel . --no-deps --no-build-isolation --wheel-dir dist- Distributed under the MIT License.
- This repository does NOT contain or bundle any original game binary data or multimedia assets. The
syncfunctionality relies on public Master Data repositories; users are responsible for confirming data source licenses and game Terms of Service before production use.