Skip to content

MCP Server

The MCP (Model Context Protocol) server exposes AgentPool tools, resources, and prompts to external MCP clients, enabling other agents and applications to use your capabilities.

What is MCP?

MCP is a standardized protocol for AI tool integration:

  • Tools - Functions that agents can call
  • Resources - Data and files agents can read
  • Prompts - Reusable prompt templates
  • Sampling - Request completions from the client

Quick Start

# Expose tools from a pool
agentpool serve-mcp config.yml

# Or expose specific toolsets
agentpool serve-mcp config.yml --toolset file_access --toolset search

See serve-mcp for all CLI options.

Use Cases

1. Share Tools with Claude Code

Expose your custom tools to Claude Code:

# Start MCP server
agentpool serve-mcp tools.yml --transport sse --port 3001

Configure Claude Code's mcp.json:

{
  "mcpServers": {
    "agentpool": {
      "url": "http://localhost:3001/sse"
    }
  }
}

2. Agent Composition

Let external agents use your pool's capabilities:

# config.yml - Your pool with specialized agents
agents:
  data_expert:
    type: native
    model: openai:gpt-4o
    system_prompt: "Expert in data processing"

  api_expert:
    type: native
    model: anthropic:claude-sonnet
    system_prompt: "Expert in API design"

tools:
  - type: subagent  # Exposes delegate_to, ask_agent

External agents can now delegate to your experts via MCP.

3. Share Custom Tools

Expose custom Python functions:

from agentpool import Agent
from agentpool.tools import tool

@tool
def analyze_data(data: str) -> str:
    """Analyze the provided data."""
    # Your analysis logic
    return f"Analysis result: {result}"

agent = Agent("analyzer", model="...", tools=[analyze_data])

agentpool serve-mcp --config agent.yml

Exposed Capabilities

Tools

All enabled tools from configured toolsets:

tools:
  - type: file_access    # read, write, list_directory
  - type: process_management      # run_command, run_python
  - type: search         # web_search, news_search
  - type: subagent       # delegate_to, ask_agent, list_nodes

Resources

Resources from configured resource providers:

resources:
  - type: file
    path: /data
    pattern: "*.json"

Prompts

MCP prompts from connected MCP servers are re-exposed:

mcp_servers:
  - "uvx prompt-server"  # Prompts become available

Architecture

graph LR
    subgraph Clients["MCP Clients"]
        Claude["Claude Code"]
        Other["Other Agents"]
    end

    subgraph Server["AgentPool MCP Server"]
        MCP["MCP Protocol"]
        Tools["Tool Manager"]
        Resources["Resources"]
    end

    subgraph Pool["Agent Pool"]
        Agents["Agents"]
        Toolsets["Toolsets"]
    end

    Claude <-->|MCP| MCP
    Other <-->|MCP| MCP
    MCP --> Tools
    MCP --> Resources
    Tools --> Toolsets
    Resources --> Pool

Transport Types

stdio (Default)

For subprocess communication:

agentpool serve-mcp config.yml --transport stdio

Used when the client spawns the server as a subprocess.

SSE (Server-Sent Events)

For HTTP-based communication:

agentpool serve-mcp config.yml --transport sse --port 3001

Consuming MCP Servers

AgentPool can also consume MCP servers (act as client):

agents:
  my_agent:
    mcp_servers:
      - type: stdio
        command: uvx
        args: ["some-mcp-server"]

      - type: streamable-http
        url: http://localhost:3001/mcp

See MCP Server Integration for consuming MCP servers.

See Also