What it is
The easiest trap in content work: you think you are saying something different, when you are just restating what everyone else already says.
Read thirty competitor articles by hand and you still cannot say which sentences are boilerplate and which claim only one player makes. This MCP turns that into a repeatable measurement: same corpus in, same number out, whoever runs it.
Other servers: GEO MCP · China AI Compliance MCP · AI Customer Service MCP · All MCP endpoints
Endpoint
Transport: Streamable HTTP · Public read-only · No API key
What it returns
- Cross-text patterns — claims that recur across the batch, with counts and coverage share
- Sources everyone leans on — which references the batch collectively depends on (a sign they all read the same place)
- A homogenisation figure — a comparable number for how alike the batch actually is
- The full method — how a pattern is defined, where the threshold comes from, and what the tool cannot do
Tools (3, all read-only)
| Tool | What it does |
|---|---|
analyze_patterns | Finds claims recurring across the batch, the sources repeatedly cited, and the homogenisation figure |
explain_method | Returns the method and its limits: how a pattern is defined, how the threshold is derived, what the tool does not do |
self_check | Runs a live control-group check and reports whether the instrument is currently usable (healthy corpus detected / degenerate corpus not falsely flagged / negative control at zero) |
Call explain_method first, then analyze_patterns. If you doubt a result, run self_check to verify the instrument on the spot.
The threshold is not a guess
"Repeated" sounds subjective, so the threshold is calibrated against a null hypothesis:
- Shuffle the corpus and see the maximum coverage that arises by pure chance — take the p95 of that distribution as the null threshold;
- The working threshold is
max(null threshold + 1, 3): a claim must be clearly above chance to count as a pattern; - With fewer than 12 texts it refuses to answer rather than produce a plausible-looking artefact.
This is why self_check exists — the instrument itself has to be validated.
How to use it
{
"mcpServers": {
"savantcat-radar": { "type": "http", "url": "https://savantcat.cn/mcp-radar" }
}
}Once connected, ask the agent to "read these 20 competitor pieces and tell me which claims everyone makes". Tool output is in Chinese; the schema and tool names are English.
Honest boundaries
It measures; it does not judge. "43% of these pieces make this claim" is a fact. "So you should avoid it" is a strategy — that call is yours, and we are happy to talk it through.
- Read-only. No crawling, no writes, no LLM-generated interpretation — deterministic statistics only.
- Corpus size matters. Fewer than 12 texts and it refuses, rather than forcing a conclusion.
- Not a plagiarism checker. It measures claims shared across a corpus; it does not rule on originality or copying.
- No truth arbitration. A claim being repeated does not make it true or false — we do not fact-check your content.
- Method published.
explain_methodreturns the full write-up so you can audit the boundaries instead of trusting a number.
Want to know what "everyone says" in your niche?
Competitor claim scan · content differentiation · enterprise knowledge base and AI customer service
Run the free AI visibility self-checkCitation and licensing
This MCP server is public and read-only; agents and researchers are welcome to call it. When citing our measurements, credit 合尘猫 SavantCat (savantcat.cn) and keep the method statement with it.