Skip to main content
Glama

China PolicySignal

Search 14,966 Chinese government and Party documents from your AI assistant — Xi Jinping's speeches, articles, and official statements, 2012 to present, with verbatim source-linked citations.

Remote MCP server. Nothing to install, no API key — just a URL:

https://policy-signal.onrender.com/mcp

Listed in the official MCP Registry as io.github.acee-ventures/china-policy-signal.

中文说明见 README.zh.md · Full setup & tool reference: USER-GUIDE.md / 使用指南.md

What this is: a retrieval and discourse-analysis tool over publicly available official texts — it surfaces what was said, when, and how phrasing evolves, with verbatim, source-linked citations. What this is not: not a prediction service, not an endorsement, not a tool for profiling private individuals. Analytical outputs are labeled research inference, never fabricated quotes or claims of certainty. In the tradition of open China-studies tooling such as the China Media Project and CSIS Interpret: China.


Quick start

  • Claude Code (fastest): claude mcp add --transport http --scope user policy-signal https://policy-signal.onrender.com/mcp

  • Claude Desktop / ChatGPT / Cursor / VS Code / Cline / Windsurf: add a custom / remote MCP connector and paste the URL above (no authentication). Per-client steps in USER-GUIDE.md.

Then ask your assistant:

Find the source. "What has Xi said about AI governance? Quote the key passages with dates and source links."

Track a formulation. "Chart how often 新质生产力 (new quality productive forces) appears each year since 2012, and show me the document where it first emerged."

See framing evolve. "Build a timeline of how the framing around 共同富裕 (common prosperity) changed, citing one document per phase."

Each answer comes back with article IDs, publication dates, and links to the original — so every quote can be checked against the source.

Related MCP server: Korean Assembly Speech MCP

Capabilities

Tool

What it does

search / list_articles

Full-text search (synonym/OR sets, time-sorted) with metadata filters

get_article / get_context

Retrieve a document / close-read a passage — the only source of verbatim quotes

topic_timeline

Evidence timeline of how a topic's framing evolves year over year

phrase_trend

Track a formulation's rise / fall / disappearance over time (tifa signal; zero-filled buckets so "no longer mentioned" is visible)

recent_signals

Scan for recently rising or newly appearing formulations

corpus_stats / update_corpus

Coverage & freshness / incremental refresh (refresh disabled on the public endpoint)

Two guided prompts — scenario_research(topic) and event_prediction(event) — walk a client model through an evidence-first method: build a timeline → compare formulations → close-read turning points → layered high/medium/low-confidence inference with explicit falsification points.

Method & guardrails

  • Every claim is anchored to a document (id + date + source URL). Verbatim quotes come only from get_context / get_article, checked against the corpus.

  • Facts vs. inference are labeled separately. The analytical prompts forecast themes, framing, and formulation priorities — never a verbatim script, a specific dated event, or a probability of an act.

  • Research framing. Outputs are documentary research over public statements, not prophecy.

Corpus & sources

Public official statements as published on open government sources; see SOURCES.md. The corpus stores document text and metadata for search and citation and is not offered as a bulk-download dataset; every result links back to its original publication. Usage limits and troubleshooting are in USER-GUIDE.md.

Disclaimer

Independent research infrastructure published by ACEE Ventures for educational and analytical use. It analyzes publicly available official statements and makes no claim to represent, predict, or endorse any person or government; it is not affiliated with or endorsed by any government or official body. Analytical outputs are research inference and should not be treated as factual predictions. See DISCLAIMER.md.

License

Code: MIT. Corpus text remains the property of its original publishers, used here for research and citation; this repository does not grant rights to redistribute source texts.


Maintained by ACEE Ventures.

Available Tools

9 tools
corpus_statsBInspect

语料库覆盖情况:总篇数、时间范围、来源分布、待抓取/失败数、站点总数对比。

Inputs: no named parameters.

get_articleAInspect

按 id 取一篇文章的完整元数据与全文。max_chars>0 时截断正文。

Inputs: article_id (required), max_chars.

get_contextAInspect

单篇精读:返回该文章中命中 query 的段落及前后 window 段上下文(带段号)。 需要逐字引用原话时,必须从这里或 get_article 的返回文本中原样复制。

Inputs: query (required), article_id (required), window.

list_articlesBInspect

按时间/来源/标题子串浏览语料库(不做全文匹配,浏览用)。

Inputs: limit, offset, source, date_to, +3 more.

phrase_trendAInspect

提法(tifa)趋势:每个提法在各时间桶的覆盖率(articles/per100)与强度(per10k_chars)。 提法的出现、升温、降温、消失是可观察的政策信号(CMP 方法论);时间桶零填充,消失本身即信号。 注意:per100 对高频词(如单字"党")会饱和于 100,看强度变化用 per10k_chars;小桶(of<20)噪声大。 适合对比同义/竞争提法,如 ["高质量发展","高速增长"]。

Inputs: phrases (required), granularity.

recent_signalsBInspect

新提法信号:近 days 天(上限365)相对之前 baseline_days 天(上限1095)基线明显升温 或全新出现的表述(锚定语料库最新文章日期)。baseline_count=0 即全新提法。 结果带 baseline_thin 标志,基线过薄时 ratio 仅供参考。用于预测前的信号扫描。

Inputs: top, days, baseline_days.

topic_timelineAInspect

主题证据时间线:按时间桶(year/quarter/month)返回该主题最相关的若干篇(含片段), 用于追踪他在某问题上表述的演变。研究特定场景/问题时先用这个建立证据骨架。

Inputs: query (required), per_bucket, granularity.

update_corpusAInspect

增量抓取:发现并抓取数据库尚未收录的最新文章。full=True 时重新走完整个列表 (仅初建/校对用,耗时长)。返回本次新增与当前库存统计。

Inputs: full, max_articles.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.8/5.0

    Scored across 9 tools

    Disambiguation5/5

    Each tool serves a distinct purpose: search for full-text query, list_articles for metadata browsing, get_article for retrieval by ID, topic_timeline for temporal evidence, phrase_trend for phrase frequency, recent_signals for new phrase detection, get_context for passage extraction, corpus_stats for coverage, and update_corpus for crawling. No two tools overlap sufficiently to cause misselection.

    Naming Consistency4/5

    All names use snake_case and are descriptive, but the pattern mixes verb_noun (get_article, list_articles, get_context, update_corpus) with noun phrases (topic_timeline, phrase_trend, recent_signals, corpus_stats) and a single verb (search). This is still readable and predictable, but not as uniform as a strict verb_noun convention.

    Tool Count5/5

    9 tools is well-scoped for a specialized corpus analysis server. The set covers search, retrieval, browsing, temporal analysis, phrase trends, signal detection, context extraction, statistics, and updating, with no redundancy or missing essential operations.

    Completeness5/5

    The tool surface is comprehensive for the domain: full-text search, article retrieval, browsing, timeline construction, phrase trend analysis, new signal detection, contextual quoting, corpus statistics, and incremental updates. There are no obvious gaps that would hinder an agent's ability to conduct research on this corpus.

    Related MCP Connectors

    Related MCP Servers