MCP + LangChain: Building AI Agents That Use MCP Tools

By /

MCP + LangChain is the 2026 power combo. MCP provides the tools. LangChain provides the agent loop (plan, act, observe, repeat). Together, you get agents that can dynamically discover and use any MCP server.

Why Combine Them?

  • MCP alone: Tools exist, but need a human to prompt the AI to use them
  • LangChain alone: Agent loop exists, but you define tools in code (not portable)
  • Together: Agent automatically discovers available MCP tools and uses them autonomously to complete complex tasks

Architecture

User: "Research competitors and create a report"
    |
    v
[LangChain Agent (ReAct loop)]
    |
    +-- Discovers MCP tools at startup
    +-- Plans: Step 1: Search web, Step 2: Read pages, Step 3: Write doc
    +-- Acts: Calls MCP tools (search, browser, filesystem)
    +-- Observes: Checks results
    +-- Repeats until done
    |
    v
Result: Competitor analysis document created

Code: LangChain Agent with MCP Tools

from langchain.agents import create_react_agent
from langchain_anthropic import ChatAnthropic
from langchain_mcp import MCPToolkit

# Connect to MCP servers
toolkit = MCPToolkit()
toolkit.add_server("search", command="npx -y @tavily/mcp-server")
toolkit.add_server("filesystem", command="npx -y @modelcontextprotocol/server-filesystem /tmp")
toolkit.add_server("github", command="npx -y @modelcontextprotocol/server-github")

# Get all tools from all connected MCP servers
tools = toolkit.get_tools()  # Automatically discovers tools from all servers

# Create agent
llm = ChatAnthropic(model="claude-sonnet-4-20250514")
agent = create_react_agent(llm, tools)

# Run
result = agent.invoke({"input": "Find the top 3 AI hackathons happening this month and save a summary to /tmp/hackathons.md"})
print(result)

What the Agent Does

  1. Discovers tools from all connected MCP servers (search, filesystem, github)
  2. Plans the task: search for hackathons, read results, format summary, save to file
  3. Executes each step by calling the appropriate MCP tool
  4. Handles errors – if search returns nothing, tries different keywords
  5. Completes – returns when the file is written

Multi-Agent with MCP

For complex tasks, use multiple agents with different MCP server access:

  • Research Agent: Has search + browser MCP access
  • Writer Agent: Has filesystem + document MCP access
  • Reviewer Agent: Has GitHub MCP access (creates PR with the content)

They chain together: Research -> Write -> Review -> Publish

Hackathon Application

This architecture is perfect for hackathon projects:

  • Build an MCP server for your hackathon’s domain (healthcare, education, agriculture)
  • Wire it into a LangChain agent
  • Demo the agent completing a multi-step task autonomously
  • Judges see both the MCP standard (infrastructure) and the agent (intelligence)

Build AI agents at hackathons on Reskilll

Scroll to Top