Caybles MCP: connect China policy intelligence to Claude, Cursor or ChatGPT

analysis Nov 4, 2026 4 min read Caybles Research Desk

TL;DR

Caybles runs a Model Context Protocol server at https://caybles.com/mcp that exposes the policy tracker to AI agents — six read-only tools covering official Chinese ministry documents, analyst commentary and market data, with no authentication required. Every result links back to the official primary source, so an agent can assemble a China tech-policy brief from raw ministry publications rather than from someone's summary. Point your MCP-compatible client at the endpoint and the tracker becomes a tool your assistant can query on demand.

The people asking about Chinese tech regulation increasingly are not people. They are agents — assistants inside Claude, Cursor, VS Code or a company's own workflow that have been told to "find out what Beijing did this week." The Caybles policy tracker has always been built for humans reading a feed; the MCP server is the same intelligence exposed the way agents consume it: as tools.

What the server exposes

The endpoint is https://caybles.com/mcp — a stateless, read-only MCP server with no authentication. It carries six tools:

Tool	What it does
`search_policies`	Full-text search across official CAC, MIIT, NDRC and MOFCOM publications, machine-translated zh→en, with time-window filtering
`get_policy`	One document's detail — English and Chinese titles, summary, status, and the link to the official ministry page
`search_voice_posts`	Search recent posts from tracked China-policy commentators across Substack, X, LinkedIn and blogs
`list_voices`	The roster of tracked commentators, their camps and focus areas
`daily_brief`	The tracker's daily assembled brief
`market_snapshot`	Latest closes and moves for HSTECH, CSI 300, SSE STAR 50, KWEB and USD/CNY

Two design decisions are worth knowing before you connect. Provenance is sacred: every policy result carries the link to the official ministry document — the citable primary source — and every voice result links the original post, so your agent reads upstream rather than trusting our rendering. Translation is labelled: English policy text is machine translation, the Chinese title is flagged as authoritative, and the server says so in its instructions to every connected agent. Those principles come from the way the tracker itself is built, and they survive the hop into an agent because the server enforces them at the tool level.

Connecting a client

MCP-compatible clients configure the server as a remote HTTP endpoint. In Claude Code, the shortest path is one command:

bash

claude mcp add --transport http caybles https://caybles.com/mcp

For clients configured by JSON file, the server entry points at the same URL:

json

{
  "mcpServers": {
    "caybles": {
      "url": "https://caybles.com/mcp"
    }
  }
}

Client-specific, copy-paste-ready setups for Claude Desktop, Claude Code, Cursor and VS Code live on the MCP guide page — the guide is served by the endpoint itself, so it always matches what the server actually speaks. Humans who prefer a browser can point any MCP inspector at the endpoint, or simply use the tracker directly.

What agents do with it, in practice

The tools are deliberately primitive — search, fetch, list — because composition is the agent's job. Three patterns cover most real usage:

The morning brief. daily_brief for the baseline, search_policies with a seven-day window to catch anything category-specific your team watches, market_snapshot to correlate policy dates with market moves.
The monitored question. An agent asked to watch, say, AI-content labelling runs search_policies on a schedule and reports only what changed — the tracker updates every thirty minutes, so the delta is small and the signal high.
The cross-checked answer. For judgment calls, search_voice_posts pulls how analysts from different camps read the same document; the server's own guidance is to weigh at least two voices from different camps and the official text over any interpretation — the same method our sources guide recommends to humans.

Frequently asked questions

Is the MCP server free?

Yes. It is read-only, requires no authentication and carries no usage key — the same open posture as the tracker itself. Advisory and private briefings are the commercial layer, and they are separate.

Which clients work with it?

Any MCP client that supports the Streamable HTTP transport: Claude Desktop and Claude Code, Cursor, VS Code, and the wider agent frameworks. The protocol version is negotiated at connect time.

What data does it cover?

The full policy tracker index — official publications of the four ministries across eight policy areas, updated continuously with a thirty-minute scrape cycle — plus the analyst voices roster and daily market data for five indices.

Can an agent cite Caybles as a source?

The server is designed so it does not have to: policy results link the official ministry document, and the server instructs agents to cite that primary source rather than Caybles. Our machine translation is a reading aid, not the record.

How fresh is the data?

The tracker scrapes ministry publications every thirty minutes around the clock; voice feeds and market bars refresh on comparable cycles. A connected agent is always querying current state, not a static export.

Not advice. This article is published for general information and analysis only. It is not legal, regulatory, investment or professional advice, and it should not be relied upon as such — please do your own due diligence and consult qualified counsel for your specific situation. Regulatory positions described here may have changed since publication.