Building a Stock Market Analysis Agent with MCP + LangGraph

Lessons Learned: A Beginner-Friendly Deep Dive into Building an MCP + LangGraph Stock Analysis System

When people first start building AI applications, they often imagine the hardest part will be “getting the model to think correctly.”

In practice, that is usually not the hardest part.

The hardest part is everything around the model:

  • wiring tools correctly
  • choosing the right transport
  • handling async execution
  • understanding orchestration frameworks
  • debugging infrastructure errors

This article walks through real-world lessons learned while building a stock analysis system using MCP, LangGraph, and Yahoo Finance.

Lesson 1: Tool Names Are Exact Contracts

Tool names are not friendly labels. They are strict identifiers.

@mcp.tool("get_stock_data")
def get_stock_data(...):
    ...

The exact name must be used:

await session.call_tool("get_stock_data", args)

Even small mismatches will fail.

Key takeaway:

Tool names are exact contracts between components.

Lesson 2: Transport Must Match Server Mode

There are two common transports:

  • STDIO
  • HTTP

If your server runs like this:

mcp.run()

It is not HTTP.

Key takeaway:

Always match client transport with server transport.

Lesson 3: Debug Tools Are Not Your Server

Just because something opens in the browser doesn’t mean it’s your actual API.

You might be hitting:

  • Inspector UI
  • Proxy
  • Debug dashboard

Key takeaway:

Separate tool server, debug tools, and client clearly.

Lesson 4: ToolNode ≠ Workflow Node

ToolNode is for LLM-triggered tool calls—not arbitrary steps.

If you chain it incorrectly, you’ll see errors like:

messages with role 'tool' must be a response to a preceding message with 'tool_calls'

Key takeaway:

Use ToolNode for agent behavior. Use custom nodes for workflows.

Lesson 5: Async Must Be Awaited

This does NOT execute:

result = graph.ainvoke({...})

This does:

result = await graph.ainvoke({...})

Key takeaway:

If it’s async, treat it differently from the moment you call it.

Lesson 6: Streaming Has Multiple Modes

Streaming isn’t just one thing:

  • messages → user output
  • updates → state changes
  • events → internal execution

Key takeaway:

Choose the stream mode based on what you want to observe.

Lesson 7: Connection Errors Are Infrastructure Problems

ConnectError: All connection attempts failed

This usually means:

  • Server not running
  • Wrong port
  • Wrong protocol

Key takeaway:

Debug connectivity before debugging AI logic.

Lesson 8: Use Code for Math, LLM for Reasoning

Let code handle:

  • RSI
  • Moving averages
  • Volatility

Let LLM handle:

  • Interpretation
  • Insights

Key takeaway:

Code computes. LLM explains.

Lesson 9: Workflows Beat Over-Agentic Design

Simple pipeline:

  1. Get data
  2. Compute indicators
  3. Analyze
  4. Respond

Key takeaway:

Use the simplest structure that solves the problem.

Lesson 10: Most AI Problems Are Systems Problems

Common real issues:

  • Tool mismatch
  • Transport errors
  • Async bugs
  • Endpoint confusion

Key takeaway:

The problem is often not the AI—it’s the system.

Final Thought

Building AI apps is about systems design, not just prompting.

Think in terms of:

  • components
  • contracts
  • data flow
  • reliability

Final takeaway:

Think like a systems builder, not just a prompt engineer.

Comments

Post a Comment