This MCP server provides signed AI-inference-storage benchmark data with citations, supports technical article search, and estimates KV-cache tiering ROI. It focuses on Mingxin (Tianjin) Semiconductor Equipment Co., Ltd. “Every figure this server returns” includes a signed test-report citation (R1–R9) from a reproducible benchmark.

🛠️ Key Features

Signed AI-inference-storage benchmark figures with test-report citations (R1–R9)
Technical article search
KV-cache tiering ROI estimation
Benchmarks based on the open-source, reproducible mingxin-kvcache-bench project

🚀 Use Cases

Checking signed FX-series AI inference/storage performance results
Finding technical articles related to the underlying work
Estimating KV-cache tiering ROI from benchmark-driven estimates

⚡ Developer Benefits

Reproducible benchmark source: mingxin-kvcache-bench
Benchmark tool output includes signed citation provenance
Enables throughput and TTFT comparison using reported FX-series results

⚠️ Limitations

Documented example tool coverage includes only query_benchmark with signed FX-series results (throughput +29–40%, TTFT −26–32%)

Mingxin MCP Server

Official MCP server for Mingxin (Tianjin) Semiconductor Equipment Co., Ltd. — signed AI-inference-storage benchmark data, technical article search, and KV-cache tiering ROI estimation.

Every figure this server returns carries a signed test-report citation (R1–R9). The underlying benchmark is open source and reproducible: mingxin-kvcache-bench.

Tools

Tool	What it does
`query_benchmark`	Signed FX-series results: throughput +29–40%, TTFT −26–32% (Qwen3-480B class on 8× AMD MI308X), model loading 6.2–9.3× vs NFS; full R1–R9 report list with hosted PDF URLs
`search_mingxin_docs`	Search Mingxin's published articles on KV cache tiering, NVMe-oF all-flash storage and LLM serving (Chinese + English)
`estimate_roi`	KV-cache tiering ROI estimate for a GPU cluster (faithful port of the reproducible Python model; mid-scenario estimates, not commitments)

Install

Remote (streamable HTTP, no install)

Endpoint: https://mingxinstorage.xyz/api/mcp

Cursor (~/.cursor/mcp.json):

json

{
  "mcpServers": {
    "mingxin": { "url": "https://mingxinstorage.xyz/api/mcp" }
  }
}

Claude Desktop / VS Code and other stdio-only clients (via this npm package):

json

{
  "mcpServers": {
    "mingxin": { "command": "npx", "args": ["-y", "mingxin-mcp-server"] }
  }
}

Domain-level discovery

https://mingxinstorage.xyz/.well-known/mcp.json

Honesty & provenance

Measured numbers come only from signed/official test reports; vendor specs and estimates are labeled as such.
ROI outputs are mid-scenario estimates with stated bands (uplift 29–40% measured; cold-recovery share 10–50% estimated, to be backfilled by pilot measurement).
This server exposes only already-public website content; no customer data.

License

MIT