Stock SDK

English

简体中文

English

简体中文

Appearance

AI Skills Overview ​

MCP tools are "atomic capabilities" — each fetches one kind of data, called on demand by the model. Skills are scenario-oriented, higher-level capabilities: they package "which tools to call, how to compute, how to summarize" into an expert-authored template, exposed through the MCP Prompts capability so a user can trigger a whole analysis pipeline with one click.

Skills are real MCP Prompts (prompts/list + prompts/get); the server declares capabilities.prompts at initialize. Prompts-aware clients (Claude Desktop, Claude Code, Cursor, Cline, …) render each skill as a slash-command / template option.

How a skill works ​

Key mental model: the server does not execute the skill. prompts/get merely hands the client's model a parameter-filled "task briefing" that names which real tools to call. The multi-step execution is done by the client-side model plus its tools/call loop.

A full analyze_stock interaction:

  1. Discover: on connect the client sends prompts/list, rendering skills as slash-commands / templates.
  2. Pick + fill: the user picks analyze_stock and fills symbol (required) and period (optional, default daily).
  3. Fetch template: the client sends prompts/get; the server interpolates the briefing and returns one user message.
  4. Inject: the briefing becomes the opening content of the user turn — the server bows out here.
  5. Model runs: the model issues tools/call in order (search → get_kline_with_indicators → get_kline_signals).
  6. Answer: the model responds in the structure the template dictates, in the user's language (every template ends with "respond in the user's language" — the instructions are English, but a Chinese question still gets a Chinese analysis).

So a skill's value isn't "unlocking capability" — it's standardized orchestration + fewer missed/mis-called tools + a one-click entry for non-experts + a uniform disclaimer and safety discipline.

Which skills are built in ​

Two tiers, core (4 by default) and full (3 advanced), filtered independently from tools:

  • core: analyze_stock (technical), screen_stocks (screening), market_overview, monitor_watchlist
  • full: analyze_capital_flow, analyze_fund, diagnose_stock

Each skill's arguments, underlying tools and example prompts are in the Skill Catalog.

Enable & trigger ​

Every MCP client uses the same entry npx -y stock-sdk mcp; only the config path and UI entry differ. For Claude Desktop (claude_desktop_config.json):

jsonc
{
  "mcpServers": {
    "stock-sdk": {
      "command": "npx",
      "args": ["-y", "stock-sdk", "mcp"],
      "env": {
        "STOCK_SDK_MCP_TOOLS": "full",     // tool tier (existing)
        "STOCK_SDK_MCP_PROMPTS": "full"    // skill tier (defaults to core)
      }
    }
  }
}
  • Claude Code: claude mcp add stock-sdk -e STOCK_SDK_MCP_PROMPTS=full -- npx -y stock-sdk mcp; skills appear as slash-commands like /mcp__stock-sdk__analyze_stock.
  • Cursor / Cline: paste the same mcpServers JSON into .cursor/mcp.json or the extension's MCP settings.

STOCK_SDK_MCP_PROMPTS accepts core (default, 4) / full (all 7) / a comma-separated skill list. Clients without Prompts support simply see only the tools — unaffected.

Full skills need the full tool set

Skill tier and tool tier filter independently. The 3 full skills (analyze_capital_flow / analyze_fund / diagnose_stock) name full-tier tools, so when you enable full skills you must also set STOCK_SDK_MCP_TOOLS=full. If you set only STOCK_SDK_MCP_PROMPTS=full while tools stay at the default core, the model hits an "Unknown tool" when its orchestration reaches a full-tier tool. The server logs a stderr warning on startup if it detects this mismatch. The example config above already pairs both — copy it as-is.

Read-only safety ​

Every skill sits on read-only tools, and each template ends with an explicit "read and analyze only — never place orders or move funds; if data is missing, say so". Compared with order-capable financial MCPs, that's a clear safety boundary.

Next steps ​