mcp-eval is built on mcp-agent — a simple, composable framework for building effective agents with Model Context Protocol using patterns from Anthropic’s Building Effective Agents guide.

​
Understanding mcp-agent

mcp-agent provides the core agent infrastructure that powers mcp-eval’s testing capabilities:
  • Simple & Composable: Build agents using proven patterns like Parallel, Router, Evaluator-Optimizer, and Swarm workflows
  • Full MCP Support: Agents can use any MCP tools, resources, and prompts from connected servers
  • Production-Ready: The same agent patterns you test can be deployed to production
  • Model-Agnostic: Works with OpenAI, Anthropic, and other LLM providers
Your test agents can leverage all of mcp-agent’s sophisticated capabilities:
  • Multi-server connections
  • Complex workflow orchestration
  • Human-in-the-loop interactions
  • Durable execution with state management
Learn more:

​
Ways to define

  • Config‑defined AgentSpec (root agents or discovered subagents)
  • Programmatic AgentSpec
  • Programmatic Agent instance
  • Programmatic AugmentedLLM
  • Factory (safe for parallel tests)
APIs: config.py, core.py

​
Decorator order

When combining, place @with_agent(...) above @task(...).

​
Discovery

Define AgentSpecs in mcp-agent.config.yaml or enable subagents search paths. Reference by name with use_agent("SpecName") or @pytest.mark.mcp_agent("SpecName"). Examples: agent_definition_examples.py

​
Per‑Agent provider/model

Prefer attaching provider/model at the AgentSpec level if an agent should use a distinct model from global defaults.