CrewAI Research Agent

Multi-agent crew with MCP web research tools.

Difficulty: Intermediate

What You'll Learn

• CrewAI @CrewBase pattern with @agent, @task, @crew decorators
• Sequential multi-agent workflow (researcher → writer)
• MCP tool integration with get_mcp_tools()
• Proxy credential substitution for LLM API keys
• Policy enforcement for MCP and HTTP access

Agents

Agent	Role	Tools
Researcher	Senior Researcher	MCP fetch tools (web fetch)
Writer	Technical Writer	None (synthesizes from research)

Setup

agentvisor template create crewai/research-agent
cd research-agent

export ANTHROPIC_API_KEY=sk-ant-...
temporal server start-dev &
agentvisor serve . --sandbox=none

Test It

THREAD=$(curl -sX POST http://localhost:8090/threads | jq -r '.thread_id')

# Research any topic
curl -sX POST "http://localhost:8090/threads/$THREAD/runs?wait=120s" \
    -H "Content-Type: application/json" \
    -d '{"input": {"topic": "the Rust programming language"}}' \
    | jq '.output.output'

# Try another topic
curl -sX POST "http://localhost:8090/threads/$THREAD/runs?wait=120s" \
    -H "Content-Type: application/json" \
    -d '{"input": {"topic": "quantum computing recent developments 2024"}}' \
    | jq '.output.output'

The crew accepts a topic input and returns a structured markdown report.

How It Works

1. Discovery: AgentVisor detects crewai.yaml and selects the CrewAI framework provider
2. Execution: python -m agentvisor.crewai.runner loads the research crew from crew.py
3. Research phase: The Researcher agent calls MCP fetch tools → AgentVisor checks policy → host calls fetch server → web content returned
4. Writing phase: The Writer agent synthesizes the research notes into a markdown report
5. Output: The final report is returned as the crew output

Credential Flow

Guest Agent (ANTHROPIC_API_KEY=mav-tok-xxx)
    └─► Guest Proxy (TLS termination)
           └─► Host Proxy (token substitution: mav-tok-xxx → real key)
                  └─► api.anthropic.com (Authorization: Bearer real-key)

The real ANTHROPIC_API_KEY is never visible inside the guest sandbox.

Policy Model

Resource	MRN Pattern	Policy
MCP fetch tools	mrn:agentvisor:mcp:fetch/.*	Allow (authenticated)
Anthropic API	mrn:agentvisor:http:api\.anthropic\.com.*	Allow (authenticated)
All other HTTP	mrn:agentvisor:http:.*	Deny

Project Files

research-agent/
├── crew.py                # ResearchCrew: researcher + writer agents
├── crewai.yaml            # Crew discovery config (auto-detected by AgentVisor)
├── mav-agent-config.yaml  # A2A metadata
├── requirements.txt       # Python dependencies
├── agentvisor.yaml        # Runtime config (proxy creds + MCP server)
└── policies/
    └── domain.yml         # Authorization policies

crewai.yaml

crews:
  research: "./crew.py:ResearchCrew"
dependencies:
  - "."

agentvisor.yaml (key sections)

proxy:
  credentials:
    - name: anthropic-key
      guest_env_var: ANTHROPIC_API_KEY
      destinations:
        - "api\\.anthropic\\.com"
      resolver:
        type: bearer_token
        source: env
        env_var: ANTHROPIC_API_KEY

mcp:
  servers:
    - name: fetch
      transport: stdio
      command: ["uvx", "mcp-server-fetch"]

Customizing the Model

export ANTHROPIC_MODEL=claude-sonnet-4-5-20250514
agentvisor serve . --sandbox=none

Default: claude-haiku-4-5-20251001

Next Steps

• CrewAI Agents Guide: Deep dive into CrewAI integration
• MCP Gateway: Configure additional MCP tool servers
• LangGraph Research Agent: Same pattern with LangGraph