I Turned the AI Customer Service National Standard into an MCP Tool, So an Agent Can Ask and Get an Answer

Last month a friend who builds AI customer service came to me with a very concrete question:

“Against this national standard, do I actually count as compliant?”

I could not answer him at the time either. Because this standard — GB/T 47746—2026, "Customer contact service — Requirements for collaboration between human and intelligent customer service" — runs to more than 40 pages with an extremely dense set of clauses, and the question of "what counts as compliant for AI customer service" is scattered from Chapter 4 through Chapter 9.

Flipping through it page by page by hand is painfully slow. And there is a good chance you finish still unsure.

So I built something: I turned the compliance Q&A for this standard into a tool an Agent can call directly.

Online address (public, read-only, no install, no key): https://savantcat.cn/en/mcp/

>

Source repository: the project mcp-savantcat on Gitee

First, how much this standard weighs

This is China's first national standard focused on "the collaboration mechanism between human agents and intelligent customer service".

Its core requirement is quite counter-intuitive: AI customer service cannot be judged only on whether it answers correctly; it must also be judged on whether it hands over to a human when it cannot answer. The standard names 5 categories of scenario that are explicitly required to hand over to a human automatically.

For any company running AI customer service, this is not a question of "should we do it" — it is something you cannot route around.

Why build it as an MCP rather than a web page

This was the key design choice.

As a web page, you would still have to read it page by page, find the clauses yourself, copy them yourself. It would only turn a paper standard into an electronic one, without solving "how should I ask".

As an MCP it is different — the Agent can ask by itself.

Any MCP-capable client (Claude Desktop, Cursor, Cline, or an Agent you built yourself) connects, and then you can simply say:

• "How does AI customer service pass the national standard?"
• "Which 5 categories of scenario must hand over to a human automatically?"
• "Where do the 61 self-check items come from?"
• "Does going live require algorithm filing or registration?"

Every answer cites the standard clause it rests on. It is not "what the AI says", it is "what chapter and article of the standard says".

For compliance work that is a hard requirement — with a sourceless "it should be fine", you cannot account to a reviewer.

The five tools

On the data side there are three files under data/: the full text of 12 atomic Q&A entries, a lightweight index, and cluster metadata. Every Q&A entry carries a sources field marking where its clause comes from.

How to connect

Claude Desktop, edit claude_desktop_config.json:

{
  "mcpServers": {
    "savantcat-answers": {
      "url": "https://savantcat.cn/mcp"
    }
  }
}

Cursor, edit .cursor/mcp.json — same content.

Hermes Agent:

mcp_servers:
  savantcat-answers:
    connect_timeout: 20
    enabled: true
    url: https://savantcat.cn/mcp

Any MCP client — a standard streamable-http handshake is enough:

curl -X POST https://savantcat.cn/mcp \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize",
       "params":{"protocolVersion":"2025-06-18","capabilities":{},
                 "clientInfo":{"name":"my-agent","version":"1.0"}}}'

No API key. Public and read-only.

Want to swap in your own industry's content? Self-host it

The architecture of this MCP separates data from code. Replace the JSON under data/ with your own content and it becomes a "compliance Q&A MCP for your industry".

pip install -r requirements.txt
python server.py
python server.py --transport http --host 0.0.0.0 --port 8765 --stateless
python server.py --selftest

There is a Dockerfile too — three commands and it is up.

One pitfall you are bound to hit when deploying

I hit it, so I am writing it down to save you time:

The MCP SDK enables DNS-rebinding protection by default and only allows localhost. Accessing it through a real domain gets you 421 Invalid Host header.

Many people's first reaction is to switch the protection off — do not do that.

The right move is to add your domain to the allow-list, keeping the protection. This project does it through TransportSecuritySettings; see DEFAULT_ALLOWED_HOSTS in server.py, and change it to your own domain before deploying.

When you hit a restriction, first think "how do I solve this inside the rules" instead of tearing out the security mechanism.

Two last words

Back to my friend's question at the start. My answer now is:

You do not need to memorise the standard — let the Agent remember it.

The standard is fixed, the clauses are dense, and human memory is limited. But an MCP tool sitting there answers whenever you ask, each answer carrying its source.

The value of this is not the time saved flipping through documents. It is that you can finally say "we are compliant" with confidence.

The standard is already in force and the tool is already live. Public, read-only, no key — connect directly:

https://savantcat.cn/en/mcp/

The source code carries the MIT licence and can be self-hosted to become a standard for your own industry (the project is mcp-savantcat, published on Gitee).

My other open-source tools are collected in the same portfolio.

See you next time.

#AI customer service #national standard #MCP #AI Agent #corporate compliance

Are you building AI customer service? Have you self-checked it against the national standard? Tell us in the comments where you are stuck.

>

If this was useful, hit like or share so more of the people who need it see it.