mcp-agent

An MCPApp holds configuration (MCP servers, model providers, logging) read from mcp_agent.config.yaml and mcp_agent.secrets.yaml. An Agent has an instruction and a list of MCP server names; entering the agent opens those connections, and attach_llm() returns an AugmentedLLM that can call the servers' tools in a loop.

Multi-agent behaviour comes from pattern helpers that each return another AugmentedLLM, so patterns nest: a router can send work to an orchestrator, an orchestrator step can fan out in parallel, and any result can pass through an evaluator-optimizer loop. The orchestrator-workers helper plans steps (all at once or one at a time), assigns them to worker agents and synthesizes the answer; a Swarm helper provides handoffs in the style of OpenAI Swarm. Control flow between patterns is ordinary Python (if, while) rather than a graph definition.

Each AugmentedLLM keeps conversation history in memory (SimpleMemory by default), which can be read, cleared or replaced. With the default asyncio engine, state lives in the process; with the Temporal engine, workflow state is persisted so runs can pause, resume and survive restarts. Long-term memory is not built in and is an open feature request.

Workflows can ask a person for input with context.request_human_input(HumanInputRequest(...)); locally this can be answered through a console callback, and the README describes human input as a tool call an agent can make. With the Temporal engine the run waits durably until it is resumed, for example with mcp-agent cloud workflows resume ... --payload '{"content": "approve"}'. MCP elicitation also lets a server pause a tool call to request structured input from the user.

pip install mcp-agent

import asyncio

from mcp_agent.app import MCPApp
from mcp_agent.workflows.factory import AgentSpec, create_orchestrator

# mcp_agent.config.yaml must define a "fetch" MCP server; set OPENAI_API_KEY.
app = MCPApp(name="brief_writer")

async def main():
    async with app.run() as running_app:
        orchestrator = create_orchestrator(
            available_agents=[
                AgentSpec(name="reader", instruction="Fetch pages and pull out facts.", server_names=["fetch"]),
                AgentSpec(name="writer", instruction="Write a five-bullet brief from the facts."),
            ],
            plan_type="full", provider="openai", context=running_app.context,
        )
        print(await orchestrator.generate_str("Summarize https://modelcontextprotocol.io for a newcomer."))

asyncio.run(main())

• Targets Python 3.10 or newer.
• Model providers are extras: pip install "mcp-agent[openai]" (also anthropic, google, azure, bedrock).
• MCP servers and providers are configured in mcp_agent.config.yaml; keys go in mcp_agent.secrets.yaml (keep it out of git) or environment variables. uvx mcp-agent init scaffolds both.
• Pause/resume, durable history and human input across restarts need execution_engine: temporal plus a running Temporal worker.
• The GitHub Releases page still shows v0.0.21 (May 2025) as latest, while tags and pyproject.toml are at 0.2.6; check the version you actually install.
• The hosted mcp-c runtime is in open beta.

Yes, it is built around MCP: agents connect to MCP servers as clients with tools, resources, prompts, sampling and elicitation, and an app can be exposed as an MCP server. A2A and AG-UI support were not found.

The repository's last push was on 25 January 2026, and nothing has been pushed since as of 30 September 2026. No deprecation notice was found in the README.

The framework is Apache-2.0 licensed. LastMile's hosted runtime, mcp-c, is described as open beta and free to use; model calls are billed by your provider.

Yes, as an MCP server: the CLI command mcp-agent install --client claude_code registers a deployed mcp-agent server in Claude Code's configuration.

mcp-agent includes a handoff pattern that its docs describe as compatible with OpenAI Swarm, but adds MCP server management, other workflow patterns such as router and orchestrator, and optional Temporal durability.