mcp-agent

mcp-agent is an open-source Python framework from LastMile AI for building AI agents that speak the Model Context Protocol (MCP) natively. Rather than inventing yet another proprietary tool abstraction, it treats MCP as the shared substrate: any MCP server you already have (filesystem, GitHub, Slack, Postgres, browser, custom) becomes an immediately usable tool for your agent, with tools, resources, prompts, notifications, OAuth, sampling, and elicitation all supported end-to-end. On top of that base, the library ships composable implementations of the workflow patterns from Anthropic's 'Building Effective Agents' essay — parallel/map-reduce, router, intent classifier, orchestrator-worker, deep-research, evaluator-optimizer, and multi-agent swarm — so you can mix and match rather than adopt one monolithic architecture. The same agent code runs on plain asyncio for local development or on Temporal for durable, pause/resume, long-running workflows without changing the API. Provider support spans OpenAI, Anthropic, Google, Azure, and AWS Bedrock, and every agent can be re-exposed as its own MCP server so other MCP-aware clients (Claude Desktop, Cursor, Windsurf, custom apps) can call into it. Configuration lives in mcp_agent.config.yaml and mcp_agent.secrets.yaml, and OpenTelemetry tracing plus token accounting are built in for production observability. Typical users are Python developers and applied-AI teams who want a lightweight, code-first alternative to LangGraph or CrewAI that leans on the MCP ecosystem instead of custom connectors, and who value being able to run the same agent locally, in a container, or on a durable execution engine without rewrites.