adata

Agent-Native financial data infrastructure for A-shares.

adata is designed for AI agents, not humans. Agents say what data they need — adata handles where to get it, how to cache it, and when to refresh it. Zero configuration, zero data-source knowledge required.

from adata import DataClient

client = DataClient()
df = client.query_daily(["sh.600519", "sz.000001"], "2024-01-01")
# Auto: check cache → fetch missing from best available provider → cache → return DataFrame

Architecture

LLM Agent
    ↓ MCP tools (query_daily / query_benchmark / data_status ...)
adata MCP Server
    ↓
DataClient — smart routing layer
    ↓ cache-first → provider fallback chain
┌──────────┬──────────┬──────────┬──────────┬──────────┐
│ Parquet   │ PolarDB  │ rqdatac  │ baostock │ akshare  │
│ (cache)   │          │          │          │          │
└──────────┴──────────┴──────────┴──────────┴──────────┘

DataClient is the single entry point. It transparently:

• Checks local Parquet cache first
• Falls back through a configurable provider chain
• Caches fetched data for future queries
• Normalizes codes across formats (sh.600519 ↔ 600519.XSHG ↔ sh600519)

MCP Server exposes DataClient as MCP tools so any LLM agent can discover and query data through the standard MCP protocol.

Installation

pip install -e .

# With specific data providers:
pip install -e ".[rqdatac]"    # 米筐
pip install -e ".[baostock]"   # baostock
pip install -e ".[akshare]"    # 东方财富
pip install -e ".[polardb]"    # Aliyun PolarDB
pip install -e ".[mcp]"        # MCP server

# Everything:
pip install -e ".[all]"

Usage

Python API (for agents and applications)

from adata import DataClient

client = DataClient()

# Daily OHLCV — auto-routed, auto-cached
df = client.query_daily(["sh.600519"], "2024-01-01")

# Benchmark index
bm = client.query_benchmark("hs300", "2024-01-01")

# Index constituents
codes = client.query_universe("hs300")

# Cache status
status = client.data_status()
# → {"data_dir": "~/.adata", "categories": {"stocks": {"files": 300, ...}}, "providers": [...]}

# Data freshness check
freshness = client.data_freshness(["sh.600519", "sz.000001"])
# → [{"code": "sh.600519", "last_date": "2026-05-14", "stale": false}, ...]

MCP Server (for LLM agents)

python -m adata mcp    # Start stdio MCP server

MCP tools available:

Tool	Description
query_daily	Query daily OHLCV data
query_benchmark	Query benchmark index (hs300, zz500, etc.)
query_universe	List index constituents
data_status	Cache overview: files, date ranges, providers
data_freshness	Check data staleness
update_data	Trigger data update by universe or codes

Claude Desktop / Claude Code config:

{
  "mcpServers": {
    "adata": {
      "command": "python",
      "args": ["-m", "adata", "mcp"]
    }
  }
}

CLI (for humans)

# Fetch data from a specific provider
python -m adata fetch --source rqdatac --universe hs300 --mode incremental

# Fetch specific codes
python -m adata fetch --source baostock --codes sh.600519,sz.000001

# Fetch benchmarks
python -m adata fetch --source rqdatac --benchmark hs300,zz500

# Full re-download
python -m adata fetch --source akshare --codes sh.600519 --mode full --start 2015-01-01

# Check cache status
python -m adata status

Data Providers

Provider	Asset Types	Auth Required	Notes
rqdatac	Stock, ETF, Index	Yes (RQDATAC_USERNAME/PASSWORD)	Fastest, most complete
tushare	Stock, Index	Yes (TUSHARE_TOKEN)	Tushare Pro, batch API
baostock	Stock	No	Free, sequential per-stock
akshare	Stock, ETF	No	Free, East Money data
polardb	Stock, Index	Yes (POLARDB_PASSWORD)	Private Aliyun DB

Provider priority is configurable:

client = DataClient(provider_priority=["rqdatac", "tushare", "baostock", "akshare"])

Default: ["polardb", "rqdatac", "tushare", "baostock", "akshare"]. Unavailable providers are automatically skipped.

Configuration

Environment variables (or .env file):

# Data directory (default: ~/.adata)
# ADATA_DIR=/path/to/data

# rqdatac
RQDATAC_USERNAME=
RQDATAC_PASSWORD=

# tushare pro (https://tushare.pro)
TUSHARE_TOKEN=

# PolarDB (private)
POLARDB_HOST=
POLARDB_PASSWORD=

Data Schema

Daily OHLCV (trade_date, stock_code, open, high, low, close, volume, amount, pct_change):

All providers normalize to this schema. Stored as per-instrument Parquet files under ~/.adata/{category}/.

Code format: {exchange}.{number} — e.g., sh.600519, sz.000001. Automatic normalization from rqdatac (600519.XSHG), PolarDB (sh600519), and raw (600519) formats.

Testing

pip install pytest
python -m pytest tests/ -v

94 tests covering all modules. All tests use in-memory fixtures, no network calls.

License

MIT