MCP endpoints: let AI call what we build

What MCP is, and why we publish endpoints

The Model Context Protocol is the open way for an AI client to call a tool instead of guessing. Without it, an agent that wants to know something about China's AI customer-service standard has to crawl pages, clean them, chunk them and hope the index is current. With it, the agent calls one endpoint and gets a structured answer with its source basis attached.

We publish these servers for two reasons. First, the material they serve is the same material we make our living from, and it is more useful callable than quoted. Second, it is a test: if the answers are wrong, anyone can point at the clause and say so — and an answer base that survives that is worth selling to a client.

Being callable is one step past being crawlable. A website can be read by AI; a service can be used by an agent. The difference matters most for facts that must be exact — standard clauses, deadlines, handover rules — where a paraphrased web page is worse than a direct query.

The five hosted endpoints

Two more, open source and running locally

These two are fully open-source MCP servers (pure Python standard library, zero third-party dependencies): clone and wire them into your client, or self-host them as remote endpoints. They are not endpoints we host for you — we store none of your data — and both the code and the method are published for audit.

Both are published to the official MCP Registry and carry an M8ven Verified badge. To see one working before committing, call simulate_dream with any text — it runs the full cycle in a temporary store that leaves nothing behind.

The thirty-four hosted tools, by name

All five servers: Streamable HTTP, public read-only, no API key, CORS open. Listed in the official MCP Registry as cn.savantcat/answers, cn.savantcat/geo-cn and cn.savantcat/ai-compliance. Answer text is in Chinese; tool names and schemas are English.

Which of the seven do you need

Connect to all five hosted endpoints at once

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

All five are Streamable HTTP · public read-only · no API key. Drop the block into your client's MCP config, then ask the agent a question in plain language.
The two open-source servers (content-router / memory-dream-engine) run locally over stdio: {"command": "python", "args": ["server.py"]} — see their pages.

What to ask it

• "Which scenarios require transfer to a human under GB/T 47746—2026, and which clause says so?"
• "Give me the compliance self-check list and mark the veto items."
• "Audit my domain for AI citability and rank the fixes by weight."
• "I asked five AI assistants about our category — are we in the source pool?"
• "We are launching an AI Q&A feature — do we need to file, and what materials are required?"
• "Read these 20 competitor articles and tell me which claims everyone makes."

House rules, limits and what is deliberately excluded

• Read-only, stateless. No tool writes to your systems, stores anything about your session, or acts on your behalf.
• No personal data, no client data. Everything served is either the published standard material or our own public documentation.
• Not the standard, and not a certificate. Clause numbers and wording are the shortest excerpts needed to locate a clause; GB/T 47746—2026 is a recommended national standard whose text belongs to its publisher, and the official published version governs. Nothing returned is an audit conclusion or legal advice.
• No scores presented as guarantees. The GEO server returns observations with timestamps; platform behaviour changes and every result has to be re-measured rather than trusted indefinitely. Where a number cannot be obtained, it is left blank with the reason stated.
• Rate limits. 60 requests per minute. If you need more, talk to us — an abused endpoint helps nobody.

Why public

Knowledge is only worth something when it is used. The number of times you get cited is the scale of trust.

Our business is putting this to work for your company: enterprise knowledge base, AI customer-service compliance, and website AI visibility (GEO). The public endpoints are the part you can test before hiring anyone — and, frankly, the part that shows whether we know the material.

Frequently asked

Do I need to install anything or pay for a key?

No. Add the config block, restart the client, done. There is no signup, no key and no quota to buy.

Can I wire these into a product I sell?

You can call them freely from an agent. To embed the underlying checklist data in your own product, use the open data files published with the self-check page and keep the attribution.

What if I need this on my own documents instead?

That is a delivery project, not an endpoint: we inventory your material, build the knowledge base and configure the same citation discipline on your side. How the knowledge base works →

Why is the answer text in Chinese?

Because the standard, the regulations and the staff who use it are Chinese. Tool names, parameters and docs are English so that any agent can drive them.