Using MCP to Do A-Share Quant Research with an AI Agent

MCP or REST? Split by scenario

Step 1: Connect the data with one command

Claude Code shown below. Config paths and key-name differences for the other 18 clients (Cursor / VS Code / Codex / CodeBuddy / WorkBuddy / Trae ...) are on the [MCP page](/en/mcp) — each checked against official docs.

claude mcp add --transport http ashareapi https://api.ashareapi.com/mcp

# Free tools need no key; for paid tools add one header:
#   "headers": { "Authorization": "Bearer <your key>" }

{
  "mcpServers": {
    "ashareapi": {
      "type": "http",
      "url": "https://api.ashareapi.com/mcp",
      "headers": { "Authorization": "Bearer ct-your-key" }
    }
  }
}

{
  "mcpServers": {
    "ashareapi": {
      "url": "https://api.ashareapi.com/mcp",
      "headers": { "Authorization": "Bearer ct-your-key" }
    }
  }
}

[mcp_servers.ashareapi]
url = "https://api.ashareapi.com/mcp"
http_headers = { Authorization = "Bearer ct-your-key" }

Step 2: Give the AI Agent work it is actually good at

Models are good at bulk, structured, repetitive work — not at "predicting price". These are the things worth handing over:

Using the tools, pull the last 60 daily bars for sh600667 and compute:
1) Close deviation from MA20 (%)
2) Last 5-day turnover vs the prior 20-day average
3) List dates where deviation and volume both expanded
Data and calculation only -- no buy or sell advice.

• Bulk screening: list the constituents of five concept sectors with their change, flag the top 10 by main-force net inflow
• Reading filings: compress three announcements into five bullet points, marking the direction of impact
• Indicator maths: pull the last 60 daily bars, compute MA20 and volume ratio, describe the position (no forecast)
• Cross-checking: align two sources and point out mismatched fields and likely causes

Step 3: Write the backtest yourself (the AI does not validate for you)

"Let the AI write the backtest code" is fine; "let the AI tell you the strategy works" is not — it does not know your cost assumptions, sample boundaries or bias sources. A minimal skeleton:

import requests

def kline(code, n=250):
    r = requests.get("https://api.ashareapi.com/v1/kline",
                     params={"code": code, "period": "day", "count": n}, timeout=10)
    return list(reversed(r.json()["data"]))   # ascending by date (API returns newest first)

rows = kline("sh600667")
# 1) Write the rule (e.g. close > MA20 and volume ratio > 1.5, hold next day)
# 2) Compute net returns: subtract round-trip cost (10cm about -0.50% / 20cm about -1.10%)
# 3) Split: define on the first 70%, look at the last 30% only (discard if direction flips)

Step 4: Adjudicate with four rules (all four required)

• Enough samples: n >= 30 — a tiny sample of wins has no statistical meaning
• Confidence interval: 95% lower bound of the expectation above 0 (bootstrap or t-test; if it contains 0, keep accumulating)
• Out-of-sample: define on the first 70%, adjudicate on the last 30%; the direction must agree
• Net first: it must stay positive after round-trip costs to count as a strategy

Free tiers and limits (stated plainly)

• Five data endpoints are free and need no key: quote / K-line / hot list / market overview / up-down distribution
• Anonymous rate limit 5/min; one PoW challenge (`GET /v1/challenge`) raises it to 60/min; keys are tiered (Standard 120/min)
• Not provided: minute bars, full-text news/announcements/reports (research reports are dehydrated summaries only), index daily K-lines — bring another vendor (e.g. Tushare); the two can coexist
• Codes need a market prefix: `sh600667` / `sz000001` / `bj8xxxxx` / `hk00700` / `usAAPL`

No data coming back? Check these five

• Wrong URL: it must be `https://api.ashareapi.com/mcp` (note the `api.` host and `/mcp` path)
• Missing prefix: `600667` will not resolve; use `sh600667`
• Shared quota: a shared egress IP hits the anonymous limit fast — add a key
• Tool list syncs on save: tools we add later require re-saving the config to appear
• Paid tools without a key: you get an explicit message, not a silent failure