kronos-mcp

Stock price forecasting for your AI agent — A-shares, US and HK equities.

An MCP server wrapping Kronos, a K-line time-series model. Ask it for a symbol, get the next N daily candles predicted: open / high / low / close / volume.

Tools

kronos_predict also returns expected_return — last predicted close over last real close, minus one. That ratio is the number most people actually want.

Quick start

You need an API key first — see below.

Claude Desktop / Cursor

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "kronos": {
      "command": "uvx",
      "args": ["kronos-mcp"],
      "env": { "KRONOS_API_KEY": "oapk_xxx" }
    }
  }
}

Command line

# no install needed
KRONOS_API_KEY=oapk_xxx uvx kronos-mcp

# or install it
pip install kronos-mcp
KRONOS_API_KEY=oapk_xxx kronos-mcp

API key

This server has no keyless mode, and cannot have one — the upstream returns 401 without a key.

The reason is cost, not gatekeeping: Kronos runs on our own GPUs and a single inference takes 30–70 seconds. Open anonymous access would let a handful of loops saturate the service, with no way to tell who is doing it.

Get one:

1. Open https://hunter.agentpit.io/dev/api-keys and sign in
2. Click 申请 API Key and choose API type KRONOS
3. Copy the oapk_… key — it is shown once
4. Set it as KRONOS_API_KEY

Approval usually lands within hours on a business day. Free, with a default quota of 1000 calls per key.

If the key is missing, this server does not return sample data or a fake success. It returns an error that tells you exactly where to get a key.

Environment variables

stdio is what Claude Desktop, Cursor and uvx use. Run a remote transport only if you want one long-running server serving several clients — and note the MCP spec marks SSE as deprecated, so prefer streamable-http for new setups.

What to expect

Every call takes 30–70 seconds. That is the GPU, not a hang. Don't retry because it feels slow, and don't loop it over dozens of symbols — upstream takes one symbol per request.

This is a statistical extrapolation, not investment advice. The model reads price history and nothing else: no news, no halts, no earnings dates. It will drift badly in unusual conditions. Treat the output as one input among many.

Errors you might hit

Related

• truesource-mcp — verifiable first-hand market signals: filings, procurement wins, macro data
• akshare-mcp — lets a model explore AKShare's 1000+ China market data functions on its own
• hunter-community — the open-source stack these come from

Apache-2.0.