Prerequisites
- Python 3.10+
- Node.js 18+
- Agent MCP Studio at agentmcp.studio
Step-by-Step Setup
Install CrewAI with MCP support
pip install "crewai[mcp]"
Start bridge.js
curl -O https://agentmcp.studio/bridge.js && npm install ws node bridge.js wss://agentmcp.studio/api/relay/YOUR-UUID
Open the studio → Settings → MCP Relay Bridge → click Connect.
-
Use MCPServerAdapter in your crew
from crewai import Agent, Task, Crew from crewai_tools import MCPServerAdapter # Load tools from your browser-based MCP server mcp_adapter = MCPServerAdapter( server_params={ "command": "node", "args": [ "/path/to/bridge.js", "wss://agentmcp.studio/api/relay/YOUR-UUID" ] } ) with mcp_adapter: tools = mcp_adapter.tools researcher = Agent( role="Data Researcher", goal="Find and analyse data using available tools", tools=tools, llm="claude-3-5-sonnet-20241022" ) task = Task( description="Query the sales database and identify the top 10 products by revenue", agent=researcher, expected_output="A structured list of top products with revenue figures" ) crew = Crew(agents=[researcher], tasks=[task]) result = crew.kickoff() print(result)
Frequently Asked Questions
Yes — CrewAI supports MCP via the MCPServerAdapter in the crewai-tools package. Connect it to your relay URL via bridge.js and any CrewAI agent in your crew can use Agent MCP Studio tools.
Use MCPServerAdapter from crewai_tools: connect to bridge.js + relay URL, call adapter.tools to get the tool list, then assign those tools to any Agent in your crew definition.
Yes — multiple agents in a crew can be assigned the same MCP tools. Each tool call routes through bridge.js → relay → your browser, so all agents access the same set of tools concurrently.
Yes — CrewAI's async execution mode works with MCP tools. The MCP protocol is inherently request-response, so async CrewAI flows can call Agent MCP Studio tools without blocking other crew operations.
CrewAI works with any LiteLLM-compatible model — OpenAI GPT-4o, Claude, Gemini, Mistral, Groq, and local Ollama models. Tool calling support depends on the model; GPT-4o and Claude have the best reliability.