Skip to main content
Guide2 min read

CrewAI MCP integration: give your crew MCP tools (2026)

CrewAI 2.1 made MCP a native part of an agent's toolset. How to connect servers with MCPServerAdapter or the new mcps field, and match tools to roles the least-privilege way.

CrewAI 2.1 made MCP a native part of an agent's toolset, so a crew of role-specialised agents can reach external systems through the same standard everything else uses. Point an agent at an MCP server and its tools join the agent's toolbox — no bespoke wrappers, no glue code. Here's the setup and the two ways CrewAI lets you connect.

Install the tools extra

MCP support ships with the CrewAI tools package. Install it alongside the core framework:

pip install crewai crewai-tools

The MCPServerAdapter pattern

The classic route is MCPServerAdapter from crewai_tools. It connects to a server over stdio, SSE or Streamable HTTP and exposes its tools as CrewAI tools you assign to any agent. The context manager keeps the connection scoped to where the tools are used:

from crewai import Agent
from crewai_tools import MCPServerAdapter

server = {"url": "https://api.githubcopilot.com/mcp/", "transport": "streamable-http"}
with MCPServerAdapter(server) as tools:
    dev = Agent(role="Engineer", tools=tools, goal="Ship the fix")

Or the mcps field

CrewAI 2.1 also added a declarative mcps field directly on an agent, so you reference servers by string for a quick start or with a structured config for full control. That keeps a crew's tool wiring in one readable place — a real help when several agents share the same servers and you want the setup version-controlled. The adapter approach is better when you need to open and close a connection tightly around a single task; the mcps field wins when the tools are a permanent part of an agent's identity.

Match servers to roles

The point of a crew is specialisation, and MCP mirrors it. Give a researcher Exa or Firecrawl, an analyst Postgres, a publisher Notion. Each agent gets only the tools its role needs — the least-privilege habit that keeps multi-agent systems safe and easy to reason about.

Going further

For orchestration design, read multi-agent orchestration patterns and role specialization. Compare the field in LangGraph MCP and Pydantic AI MCP. Vet every server with MCP security best practices.

Loadout

Build your AI agent loadout

The directory of MCP servers and AI agents that actually work. Pick the right loadout for Slack, Postgres, GitHub, Figma and 20+ integrations — with install commands ready to paste into Claude Desktop, Cursor or your own stack.

© 2026 Loadout. Built on Angular 21 SSR.