Before and After
To understand why the Model Context Protocol matters for research, consider how a cross-border analysis worked before MCP — and how it works now.
Before. A researcher wants to compare the valuation of a Chinese copper smelter listed on the Shanghai Stock Exchange with its closest US peer. The workflow: open AKShare or Wind to pull the A-share company’s financials. Open a global financial terminal for the US peer. Export both to Excel. Manually align the fiscal years (Chinese companies often report on different calendar conventions). Normalize for accounting differences between CAS and US GAAP. Build the comparison table. Copy the table into a document. Write the analysis. Elapsed time for a competent researcher: 2-4 hours. Cognitive overhead: scattered across five tools and three tabs, constant context-switching degrading analytical quality.
After. The researcher asks Claude directly: “Compare Jiangxi Copper and Freeport-McMoRan on TTM EV/EBITDA.” Claude sends parallel requests through MCP to two servers — AKShare for A-share data, global data source for US data — both queries execute concurrently, results return in the same conversation. How the data gets normalized, how accounting differences get adjusted — that is the data layer’s job. MCP’s role is giving the model simultaneous reach to both data sources, so the researcher does not have to manually shuttle data between them. Two minutes, cognitive flow unbroken.
This is not a hypothetical. This is how KSINQ’s research workflow operates today.
What MCP Actually Is
MCP — the Model Context Protocol — is an open standard introduced by Anthropic in November 2024, now governed by the Linux Foundation’s Agentic AI Foundation. Every major AI provider has adopted it. By March 2026, MCP SDK downloads exceeded 97 million across Python and TypeScript, making it the fastest-adopted AI infrastructure standard in history.
The technical description is straightforward: MCP defines a standardized interface through which an AI model can connect to external data sources and tools. Instead of building a custom connector for each data source (the “N times M problem”), you build one MCP server per data source, and any MCP-compatible AI client can connect to it.
The analogy circulating in the developer community is “USB-C for AI” — a universal connector replacing a drawer full of proprietary cables. The analogy is accurate, and the implications for research are substantial.
Why MCP Changes Research
The traditional financial data model is the terminal. You pay for access, sit in front of the screen, query manually, synthesize manually. The AI era’s first iteration was “copy-paste into ChatGPT” — a marginal improvement that preserved the bottleneck of human data retrieval.
MCP eliminates that bottleneck. With MCP, the AI does not wait for you to feed it data. It reaches out to data sources directly, queries them in real-time, and synthesizes across sources in a single analytical pass. The researcher’s role shifts from data retrieval to analytical judgment — which is where human value actually resides.
Cross-border research benefits most because data fragmentation is worst there. A cross-border researcher’s data is scattered across five or six systems that do not talk to each other (which data sources and why — see Data Layer). Before MCP, researchers spent more time assembling data than analyzing it. After MCP, assembly is automated. The researcher’s entire cognitive budget goes to analysis.
KSINQ’s MCP Architecture
We operate four MCP servers, each wrapping one category of data source:
| MCP Server | Data Source | Details |
|---|---|---|
| China Market Data | AKShare (A-shares, HK equities, futures, funds, macro) | Data Layer |
| Global Financial Data | Licensed cross-border provider (US equities, fixed income, derivatives, ESG) | Data Layer |
| Research Content | Readwise (sell-side reports, academic papers, news) | See below |
| Workflow Orchestration | Dify (pipeline scheduling, morning briefings) | Research Workflow |
Data source selection rationale, cleaning, and normalization logic belong to the data layer — not repeated here. This page covers the MCP interface: how these sources are unified and exposed to the model.
Key design decision: one MCP server per data source, not one monolithic server. The reason is isolation. AKShare’s API is occasionally unstable. If it shared a server with the global data source, one going down would take the other with it. Separated, Claude can continue analysis with global data when AKShare times out — degradation, not interruption.
Readwise: Qualitative Data via MCP
Quantitative data (prices, financials) comes through AKShare and the global data source. But research is not just numbers — what the sell-side thinks, what academia says, what the news is reporting — these qualitative signals shape judgment too.
Readwise is our qualitative ingestion layer. It automatically collects and indexes the sell-side reports, academic papers, long-form articles, and news we subscribe to. Connected via MCP, Claude can query this library directly:
- “What is the sell-side consensus on 2026 copper demand growth?”
- “Have any analysts flagged the same margin divergence I’m seeing in the data?”
- “Any regulatory developments on this company in the past two weeks?”
The Readwise MCP server does one thing: receives a query, searches the index, returns matching highlights and document summaries. It does not analyze — that is the model’s job. But it compresses “sift through three months of reading notes for relevant passages” from 30 minutes to 5 seconds.
A real usage scenario: during a cross-border valuation comparison, Claude pulled the Chinese company’s financials from AKShare, pulled the US peer’s data from the global source, and spotted a gross margin divergence. It then queried Readwise and found three sell-side reports pointing to the same explanation — a shift in Chinese government subsidy policy for that industry. The model combined all three inputs into an analysis with both quantitative backing and qualitative validation. The entire process happened in one conversation.
What MCP Does Not Do
MCP is plumbing, not intelligence. It delivers data to the model; it does not tell the model what to do with that data. Analysis quality depends entirely on the model’s reasoning capability (which is why our Claude selection matters) and the human researcher’s judgment (which is why our Triple-Perspective Framework matters).
MCP does not solve data quality problems either. If AKShare returns stale data, or the global data feed has an error in a financial statement, MCP will faithfully deliver that error to the model. Our workflow includes data validation checks, but these are downstream of MCP — part of the research process, not the connectivity layer.
MCP is infrastructure. What you build on top of it creates the research value.