mcp-agent

What it does

• Composable workflow patterns over MCP servers — map-reduce, orchestrator, evaluator-optimizer, router and more — with servers as first-class tools
• Create MCP servers with a FastMCP-compatible API, and expose agents themselves as MCP servers
• Full MCP surface: tools, resources, prompts and notifications, plus OAuth, sampling, elicitation and roots
• Durable execution on Temporal with no API changes; a beta cloud path (or the mcp-c managed runtime) deploys every app as an MCP server
• CLI scaffolding: uvx mcp-agent init generates a starter project, uvx mcp-agent deploy ships it

When to use it

Python teams building production agents on top of MCP servers who want workflow patterns and durability without hand-rolling orchestration — or anyone who wants to expose a whole agent as just another MCP server.

Getting started

# requires uv (or pip)
uv add "mcp-agent"
# or: pip install mcp-agent

# optional provider extras:
uv add "mcp-agent[openai, anthropic, google, azure, bedrock]"

# scaffold a starter project:
uvx mcp-agent init --template basic --dir my-first-agent

Things to know

Configuration is YAML-first (mcp_agent.config.yaml with an mcp.servers block) plus a separate secrets YAML — not a pure-Python wiring story. The cloud deploy path is marked beta in the README. LLM provider calls use your own keys, supplied via the secrets file or the launch environment.