AI Customer Service MCP: let agents query our answer base

What it is

A public read-only service built on the Model Context Protocol. Any MCP client — Claude Desktop, Cursor, Cline, or your own agent — can query our answer base without writing a single scraper.

It solves a concrete problem: instead of crawling pages, cleaning, chunking and indexing the material yourself, one connection gives your agent structured answers with their source basis and a compliance self-check list.

The material behind it is the same work we do for clients, published in public form: the clause-level reading of GB/T 47746—2026 (all 61 self-check items: 48 shall + 4 should + 9 may, including the 5 veto items) and the enterprise knowledge-base questions we answer every week.

Other server: GEO MCP (website AI-citability audit) · All MCP endpoints

Endpoint

Transport: Streamable HTTP (stateless) · Protocol version: 2025-11-25 · Server name: savantcat-answers
Public read-only · No API key · CORS open · Rate limit 60 req/min

Tools (5)

The same checklist is public in human-readable form: Chinese one-pager (A4 print) · English one-pager; open data as JSON / CSV, licensed CC BY 4.0 (attribution + source link); commercial use requires permission.

What an answer actually carries

• A structured body — the requirement or the answer, written to be read on its own, not a paragraph that only makes sense inside a longer article.
• Its source basis — which clause of which standard, or which class of company document, the answer rests on. An agent that quotes us can also show where the statement comes from.
• Related questions — the natural follow-ups, so a client can keep the thread instead of starting a new search.
• Plain wording in Chinese, because that is the language the standard and the client's staff work in. Tool names and schemas are English; the answer text is not.

What it does not carry: opinions dressed as requirements, client names, or any personal data. Every answer points at a public document or a standard clause, or it does not ship.

How to connect

1. Generic MCP client (JSON config)

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

Drop that block into the MCP configuration file of Claude Desktop, Cursor, Cline or any other client. On some client versions the type field is written transport: "streamable-http".

2. Plain HTTP call (any language)

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":"tools/list","params":{}}'

3. Ask a real question

curl -X POST https://savantcat.cn/mcp \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{
    "jsonrpc":"2.0","id":2,"method":"tools/call",
    "params":{"name":"search_answers",
              "arguments":{"query":"which scenarios require human handover","top_k":3}}
  }'

Listed in the official MCP Registry as cn.savantcat/answers. The answer content itself is in Chinese.

What agents do with it

• Answer compliance questions without hallucinating. A customer-service team asks an assistant whether a timeout transfer is required; the assistant calls standard_info and search_answers, then answers with the clause number instead of a guess.
• Walk through a self-check. A consultant runs self_check_list and turns the result into a gap list for a client — including the veto items that cannot be argued away.
• Ground internal documentation. A tool-builder wiring a Chinese-market customer-service product uses the answers as reference material, so the feature list follows the standard rather than a blog post about it.
• Give a stronger answer than a web search. Search returns pages; this returns the requirement, its basis and the follow-up questions in one call.

Why public

Knowledge is only worth something when it is used. The number of times you get cited is the scale of trust. Publishing the clause reading as a callable service also makes it falsifiable: if an answer is wrong, someone can point at the clause and say so, and we fix the entry.

There is also a practical reason. Most companies do not need a public service — they need their own documents to behave like this. If what you want is your product data, policies and service rules turned into a knowledge base your staff and your agents can call, with the same citation discipline, that is our main business.

Limits, and honesty about them

• Read-only. No tool writes to your system, stores anything about you, or sends messages on your behalf.
• Not the standard itself. Clause numbers and wording are the shortest excerpts needed to locate a clause. GB/T 47746—2026 is a recommended national standard and its text belongs to its publisher; the official published version governs. We are not the standard-issuing body.
• Not a compliance certificate. Nothing returned by this service is an audit conclusion or legal advice; third-party assessment under clause 9.2 must come from a body holding professional customer-service qualifications.
• Rate-limited and unauthenticated. 60 requests per minute, shared by everyone. For anything heavier, talk to us instead of hammering the endpoint.
• Chinese answer text. If you need English-language answers for your own corpus, that is a delivery question, not an endpoint setting.

Frequently asked

Do I need an account or a key?

No. The service is public and read-only; there is no signup, no key and no quota to buy.

Can I use it in a commercial product?

You can query it freely from an agent. For embedding the underlying checklist data in your own product, use the open data files published on our self-check page and keep the attribution.

Is this the same content as the free self-check tool?

Same source, different shape. The web self-check is built for a person filling in 61 radio buttons; the MCP server is the same material for an agent that needs one answer at a time.

Does an out-of-date answer get corrected?

Yes, and the version date is part of the answer set. The checklist is on version v2.0 (2026-09-19), verified against the standard text clause by clause.