Chinese History

CI PyPI License: MIT Data: CC BY 4.0 Python 3.9+ Dependencies: zero MCP Release

A traceable Chinese-history MCP server. Four Model Context Protocol tools over 9 classical Chinese texts (pre-Qin to Wei-Jin — 史记 / 汉书 / 后汉书 / 三国志 / 左传 / 论语 / 孟子 / 吕氏春秋 / 资治通鉴). Every result carries a 【book → chapter → paragraph】 citation, and honestly reports its review_status — the server never claims per-item human review it doesn't have.

一个可溯源的中国历史故事 MCP server：按事件 / 人物 / 今地名 / 品质四轴查询 先秦-汉魏九部正史子书，每条返回都带原文出处，机器生成/机审内容如实标注。

• Zero runtime dependencies — pure Python standard library. No pip install of a framework, no MCP SDK; the whole server is auditable in a few files.
• Read-only — opens the corpus with mode=ro + PRAGMA query_only; never writes.
• Honest by construction — machine-generated punctuation / translation and machine-adjudicated status are labeled in every response (AIGC-compliant).

Why this exists: as of mid-2026 the public MCP ecosystem has no classical Chinese / Chinese-history server. This fills that gap. Income expectation is zero; the goal is a useful public good.

Contents: The four tools · Install & run · The corpus database · Honesty · Data & provenance · Design notes

The four tools

Each tool call returns JSON. Multi-source events, person appraisals, and place/quality edges all carry the exact 【book → chapter → paragraph】 they came from — that is the point of the server.

Install & run

Requires Python 3.9+ (standard library only — nothing else is installed). The server speaks MCP over stdio (newline-delimited JSON-RPC 2.0).

pip install chinese-history-mcp
# then (after downloading corpus.db from Releases — see below):
chinese-history-mcp --db /path/to/corpus.db

Or run without installing, straight from a checkout:

PYTHONPATH=src python3 -m storyextractor.mcp.server --db /path/to/corpus.db

Configure in an MCP client

Claude Desktop (claude_desktop_config.json), Cline, Continue, etc. — add one stdio server. After pip install chinese-history-mcp:

{
  "mcpServers": {
    "chinese-history": {
      "command": "chinese-history-mcp",
      "args": ["--db", "/path/to/corpus.db"]
    }
  }
}

{
  "mcpServers": {
    "chinese-history": {
      "command": "python3",
      "args": ["-m", "storyextractor.mcp.server", "--db", "/path/to/corpus.db"],
      "env": { "PYTHONPATH": "src" },
      "cwd": "/absolute/path/to/chinese-history-mcp"
    }
  }
}

Try one handshake by hand

printf '%s\n' \
  '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{}}}' \
  '{"jsonrpc":"2.0","id":2,"method":"tools/list"}' \
  '{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"query_by_quality","arguments":{"quality":"忠","limit":2}}}' \
  | chinese-history-mcp --db /path/to/corpus.db

Demo + hallucination comparison

python3 scripts/mcp_demo.py --db /path/to/corpus.db runs a scripted tour of all four tools (also a minimal MCP-client reference). See docs/MCP_DEMO.md for a side-by-side of a bare LLM (fabricated / uncitable) vs. this server (cited) on the same questions.

The corpus database

corpus.db is not in this repository (it is a ~90 MB binary). Download it from this repo's Releases and point --db at it, or set STORYEXTRACTOR_DB=/path/to/corpus.db.

The database is read-only at runtime. If you host it on a read-only medium, make sure the release artifact was produced with sqlite3 corpus.db "VACUUM INTO 'corpus_release.db'" (single file, no -wal/-shm sidecars).

Honesty (please read)

This server is designed for provenance, not to launder machine output as scholarship. Downstream clients and LLMs must not present its results as "individually human-reviewed." Every response labels what it is:

• Events review_status='approved' — mostly machine bulk-approved credible inferences, not per-item human review.
• Person profiles review_status='draft' — LLM-synthesized, not human-vetted.
• Quality mappings — auto_approved = multi-LLM machine consensus, draft = pending review; evidence_quote is a real substring of the source, rationale is an LLM's reasoning.
• Place mappings — mostly multi-LLM machine consensus (auto_approved), a few human-approved; confidence is bucketed high/medium/doubtful.
• Text — original is public-domain 白文 with machine-generated punctuation/segmentation; vernacular translation is fully machine-generated.

The server also does not eliminate downstream hallucination: it gives you citable retrieval facts; an LLM built on top can still confabulate around them. The citations are anchors for human verification.

Scope is the 9 texts above — "not found" means "not in this corpus," not "did not happen."

Data & provenance

• Original text: public-domain classical Chinese 白文 (unpunctuated base text from public-domain editions), with self-produced, machine-generated punctuation and segmentation (not copied from any modern annotated/collated edition).
• Vernacular translation: machine-generated across the whole corpus.
• Annotations (events / entities / places / qualities): machine-assisted, with human review gating on selected layers; status is reported per record.

License

• Code (this repository): MIT — see LICENSE.
• Corpus data (corpus.db, distributed via Releases): CC BY 4.0.

The text layer is self-produced (punctuation/segmentation) over public-domain base text, so it is distributed freely; machine-generated attributes are labeled throughout for AIGC compliance.

Design notes

• Pure stdlib hand-written stdio JSON-RPC 2.0 (initialize / tools/list / tools/call + ping / notifications). No third-party MCP SDK.
• Read-only DB access (src/storyextractor/mcp/db.py): mode=ro + PRAGMA query_only; the migration-running db.connect is never used at serve time.
• Tests: python3 tests/test_mcp_server.py (read-only enforcement, protocol shapes/error codes, honest review_status, alias token-exact matching + disambiguation, LIKE-wildcard escaping) — builds a temporary fixture DB, so it runs without corpus.db.

Contributing & project meta

• CONTRIBUTING.md — how to run tests/lint and the principles this project holds to.
• CHANGELOG.md — release history.
• SECURITY.md — threat surface (read-only, no network) and how to report issues.

Issues and pull requests are welcome. Please keep the constraints in mind: zero runtime dependencies, read-only, every result cited, honest review_status.