Using an MCP Server as toolset for Agent which run in docker container over MCP Gateway

[ganaaduvva]

I am running an MCP Server in a container and an MCP Gateway using docker-compose
As my MCP Server is runnnig in a container with --transport=stdio
I want to use this server as toolset for my agent to use

help on how can I implement

Thanks in advance!!

[adk-bot]

Response from ADK Answering Agent (experimental, answer may be inaccurate)

TLDR: To use an MCP Server as a toolset across different Docker containers, you need to switch from a stdio transport to a network-based transport like HTTP/SSE. Your MCP Server should listen on a network port, and your agent's MCPToolset should connect to it using SseConnectionParams instead of StdioConnectionParams.

---

Hello! Based on the information in the ADK documentation, I can provide some guidance.

The --transport=stdio setting and the corresponding StdioConnectionParams in the ADK are designed for communication between processes on the same host. This allows an agent to directly start and communicate with the server's process via its standard input/output channels. [1]

When you run your MCP Server and agent in separate Docker containers, they are in isolated environments and cannot communicate directly via stdio. They need to communicate over the Docker network instead.

Here is the recommended solution:

1. Modify your MCP Server: Your server needs to be configured to listen on a network port using an HTTP/SSE transport instead of stdio. The exact implementation will depend on your server's framework, but it generally involves running a web server.

2. Modify your Agent's Toolset: In your agent's code, you should replace StdioConnectionParams with SseConnectionParams and provide the URL where the MCP Server container can be reached over the network.

Example Agent Configuration:

Here is how you would configure the MCPToolset in your ADK agent's code:

from google.adk.tools.mcp_tool.mcp_toolset import MCPToolset
from google.adk.tools.mcp_tool.mcp_session_manager import SseConnectionParams

# URL of your MCP Server container.
# If you are using docker-compose, this might be 'http://<service_name>:<port>'
# For example: 'http://mcp-server:8081'
mcp_server_url = "http://your-mcp-server-address:8081"  # Replace with your server's actual network address and port

mcp_toolset = MCPToolset(
    connection_params=SseConnectionParams(url=mcp_server_url),
)

# You can then add this toolset to your agent, for example:
# agent = LlmAgent(..., tools=[mcp_toolset])

The documentation on "MCP Tools" provides more details on connection management for different environments, including containerized setups which are conceptually similar to what you are building with docker-compose. [1]

By making these changes, your agent will be able to correctly connect to your MCP Server over the network and use it as a toolset.

[1] https://google.github.io/adk-docs/tools/mcp-tools/

[ganaaduvva]

connection_params = TCPConnectionParams(host='localhost', port=9000)

does it work then?