Serve a team of agents from one MCP server

One MCPDeployer can serve many targets. Pass a dict and every entry becomes its own tool, named by its key. A client then picks the right specialist per call instead of connecting to four servers.


Class	`MCPDeployer`
Tools	`research`, `critique`, `write_briefing`, `word_count`
Client	`MCPManager`, no agent needed
Model used here	`claude-sonnet-5`

Prerequisites

pip install -U swarms "mcp>=2.0.0"
export ANTHROPIC_API_KEY="sk-ant-..."

Build it

Define the team

Two agents, a workflow that chains two agents, and a function. Any object with a run(task) method can be served, so a swarm is served the same way as a single agent.

team.py

from swarms import Agent, MCPDeployer, SequentialWorkflow

researcher = Agent(
    agent_name="Researcher",
    agent_description="Lists the five most important facts about a topic.",
    system_prompt="List the five most important facts about the topic.",
    model_name="claude-sonnet-5",
    max_loops=1,
    print_on=False,
)

critic = Agent(
    agent_name="Critic",
    agent_description="Names the single weakest point in a piece of text.",
    system_prompt="Name the single weakest point in the text and explain why.",
    model_name="claude-sonnet-5",
    max_loops=1,
    print_on=False,
)

writer = Agent(
    agent_name="Writer",
    system_prompt="Turn the facts into a tight three-paragraph briefing.",
    model_name="claude-sonnet-5",
    max_loops=1,
    print_on=False,
)

briefing = SequentialWorkflow(
    name="Briefing-Pipeline",
    description="Researches a topic, then writes a three-paragraph briefing from the facts.",
    agents=[researcher, writer],
    max_loops=1,
)


def word_count(task: str) -> int:
    """Count the words in the text."""
    return len(task.split())


deployer = MCPDeployer(
    {
        "research": researcher,
        "critique": critic,
        "write_briefing": briefing,
        "word_count": word_count,
    },
    name="Editorial-Team",
    api_keys=["sk-local-dev"],
    port=8001,
    timeout=300,
)

Each tool’s description comes from the target: agent_description for an agent, description for the workflow, and the docstring for the function. timeout=300 fails any single call that runs longer than five minutes.

Serve it and call each tool

Add a client to the bottom of team.py. The with block starts the server on a background thread and stops it on exit, so the whole example is one script.

team.py

from swarms.tools.mcp_manager import MCPManager

if __name__ == "__main__":
    with deployer:
        client = MCPManager(mcp_url=deployer.url, api_key="sk-local-dev")

        print(client.list_tool_names())

        count = client.call_tool("word_count", {"task": "the quick brown fox"})
        print(count["result"])

        brief = client.call_tool(
            "write_briefing",
            {"task": "The history of the transistor"},
        )
        print(brief["result"])

python team.py

['research', 'critique', 'write_briefing', 'word_count']
4

followed by the briefing. word_count never touches a model, so it answers at once. write_briefing runs the researcher and then the writer before it returns.

Let an agent choose the tool

With the server running, a client agent sees all four tools and decides which to call:

from swarms import Agent
from swarms.schemas.mcp_schemas import MCPConnection

editor = Agent(
    agent_name="Editor",
    system_prompt="Use the team's tools to research, critique, and brief. Return the final briefing.",
    model_name="claude-sonnet-5",
    mcp_url=MCPConnection(url="http://127.0.0.1:8001/mcp", api_key="sk-local-dev"),
    max_loops=3,
    print_on=False,
    output_type="final",
)

To keep the server up for a separate client process, call deployer.run() instead of using the with block.

Adding tools later

add_tool registers one more target before the server starts. After start() it raises RuntimeError.

deployer.add_tool(critic, name="second_opinion", description="A second critique pass.")

Two targets that resolve to the same name raise ValueError. That is why several targets are passed as a dict: the keys make every name explicit.

Checking what is served

GET /health lists every tool:

curl http://127.0.0.1:8001/health

{"status":"ok","name":"Editorial-Team","tool":"research","tools":["research","critique","write_briefing","word_count"],"transport":"streamable-http"}

Serve an agent over MCP

The single-agent version, with a separate server and client.

MCPDeployer reference

Custom auth, token verifiers, SSE and stdio transports.

Index

Basic Examples

Model Providers

Single Agent

Multi-Agent Examples

Auto Agent Builder

Scheduled Agents

Applications

Research

Technical Analysis

Use Cases

Finance

Voice Agents

Integrations

MCP Tool Integrations

Deployment

CLI

Serve a team of agents from one MCP server

Prerequisites

Build it

Adding tools later

Checking what is served

Next

Serve an agent over MCP

MCPDeployer reference

​Prerequisites

​Build it

​Adding tools later

​Checking what is served

​Next

Serve an agent over MCP

MCPDeployer reference

Prerequisites

Build it

Adding tools later

Checking what is served

Next