How AI Agents Pull A-Share Heat Data via MCP in About Five Minutes

Most LLM agents that touch Chinese equities hit the same wall: the model can reason about a sector, but it has no structured, timely view of what the market is actually talking about. Scraping forums and news sites is brittle, slow, and full of raw text you probably should not be redistributing.

Lava (熔岩) is a read-only data service that turns A-share news flow into structured heat signals for stocks and concept boards, exposed over both MCP and REST. This post walks through what the data means, which tools exist, and how to wire an agent to it quickly.

Lava returns attention/heat data only. No bullish/bearish labels, no original text, no price targets, and nothing in this article is investment advice.

Lava counts and weights mentions of A-share stocks and concept boards across institutional notes, commentators, and mainstream financial news. Sources are anonymized into stable IDs (for example an agency ID), so you can track source quality without seeing names or source text.

Three data categories matter most for agents:

A-share narratives move through concept boards (AI computing, solid-state batteries, low-altitude economy, etc.). Lava tracks:

Comparing those across windows gives you a rotation view: attention leaving one board and arriving at another.

A first-mention is a stock that had zero mentions in the previous 7 days and just got its first one. It is a novelty signal: useful for "what entered the conversation today" digests, and a natural input for research triage.

The MCP server and the REST API share one schema. Tool names below match the OpenAPI operationId s, so an agent can resolve them deterministically.

ts_code uses the Tushare-style format, e.g. 300750.SZ, 688256.SH.

Point-in-time (PIT) scores matter if you backtest: listAgencyPit gives source scores as they were on as_of_date, so weighted heat in your backtest does not leak future information.

Sign up, copy your API key (format lvk_…), and note the MCP endpoint URL from the docs. Keys go in a header:

Authorization: Bearer lvk_xxxxxxxx

Never commit the key; read it from an environment variable.

Most MCP clients (Claude Desktop, Cursor, Cline, custom agents) accept a remote HTTP server entry like this:

{
  "mcpServers": {
    "lava": {
      "url": "<LAVA_MCP_URL>",
      "headers": {
        "Authorization": "Bearer ${LAVA_API_KEY}"
      }
    }
  }
}

Restart the client; it should list the tools above.

Call getAccount first. It costs nothing and confirms auth. Then try:

listHotConcepts(window="24h", limit=10)

If your agent framework does not speak MCP, the same data is plain HTTP:

curl -s "$LAVA_BASE/hot/first-mentions?hours=24&min_score=60&limit=20" \
  -H "Authorization: Bearer $LAVA_API_KEY"
python
import os, requests

BASE = os.environ["LAVA_BASE"]
H = {"Authorization": f"Bearer {os.environ['LAVA_API_KEY']}"}

def lava(path, **params):
    r = requests.get(f"{BASE}{path}", headers=H, params=params, timeout=20)
    r.raise_for_status()
    return r.json()

concepts = lava("/hot/concepts", window="24h", limit=10)

Goal: a short brief of "what the market is discussing", generated before the open.

listHotConcepts(window="24h", limit=10) → top boards.listConceptStocks(concept=..., days=1, limit=5). listFirstMentions(hours=24, min_score=60, limit=10) → new names.listAcceleratingHotStocks(window="24h", min_mentions=3, limit=10) → movers in attention. Prompt the model to describe attention only: "rank by heat, note rotation versus yesterday, list first-mentions". Instruct it explicitly not to produce buy/sell language.

Run listHotConcepts at window="24h" and window="7d" and diff the ranks. A concept high on 24h but low on 7d is newly rotating in; the reverse is cooling off. Store daily snapshots, or skip the bookkeeping and pull getDailyHeatPanel(type="concept", names=...) for history.

When a user asks about a ticker, have the agent fetch:

getStockHeatWindows(ts_code=...) for the 1h–7d curve;listStockConcepts(ts_code=..., days=7) for which narratives it is riding;getStockAgencyCoverage(ts_code=..., days=30, min_score=60) to see if coverage breadth is growing or shrinking. This gives the model grounded context ("heat is concentrated in the last 6 hours, mainly via the AI computing board, with coverage from more high-score sources than last week") instead of hallucinated sentiment.

min_score`` ts_code over nameslimit and avoid loops that fan out over every stock in a board. If you want to see the data before writing code, the AI board heat monitor (https://lava.ai-dolphin.com/monitor/board-ai-heat) shows a concept heat series built from these same endpoints.

An agent with Lava attached can answer "what is the A-share market paying attention to, and how is that changing?" with structured numbers instead of guesses. Setup is one header and one server entry; the useful part is combining heat, concept rotation, and first-mention into workflows.

Research notes, methodology, and worked examples are here: https://lava.ai-dolphin.com/research