Step 3: MCP (Model Context Protocol)

← Back to Index

Step 3: MCP (Model Context Protocol)

For this section you will need, you need to install Docker Desktop.

MCP is an open-source standard for connecting AI applications to external systems.

MCP is a standardised protocol clients can use to give an agent access to different MCP servers that contain tools (and much more, we encourage you to go and read the specification).

The Docker MCP Toolkit is a gateway that lets you set up, manage, and run containerized MCP servers and connect them to AI agents.

MCP Toolkit

Naturally, cagent makes it easy to use an MCP server from the MCP Toolkit.

Here is an example agent that uses the fetch MCP server from the toolkit.

version: "2"

agents:
  root:
    model: openai/gpt-4o
    instruction: Summarize URLs for users
    toolsets:
      - type: mcp
        ref: docker:fetch

Run this agent and ask it "fetch rumpl.dev", you will see the agent calling the fetch tool form the fetch MCP server.

This is the simplest way to use an MCP server with cagent, but you can also run any MCP server, local or remote.

Custom local MCP server

To run the above example you will need the uvx command line, which is often needed when running local mcp server outside of docker containers.

yfmcp is the yahoo finance mcp server.

version: "2"
agents:
  root:
    model: opeani/gpt-4o
    add_date: true
    instruction: |
      Analyse the given stock
      Use tables to display the data
    toolsets:
      - type: mcp
      command: uvx
      args: ["yfmcp"]

Remote MCP server

version: "2"

agents:
  root:
    model: anthropic/claude-3-5-sonnet-latest
    description: Moby Project Expert
    instruction: You are an AI assistant with expertise in the moby/moby project's documentation.
    toolsets:
      - type: mcp
        remote:
          url: https://gitmcp.io/moby/moby
          transport_type: streamable

Your own MCP server

In this exercice we will take an existing (very simple) MCP server, build it, and then use it in our agent.

First clone https://github.com/rumpl/mcp-strawberry

You can then build the docker image of this server:

$ git clone https://github.com/rumpl/mcp-strawberry
$ cd mcp-strawberry
$ docker build -t mcp-strawberry .

If you don't want to clone and build, you can use the already pushed MCP server image from hub: djordjelukic1639080/mcp-strawberry

Now create an agent that will use this MCP server.

Hint 1

To run any server, use

toolsets:
  - type: mcp
    command: ...
    args: [..., ...]
Hint 2

Make sure to add the -i and --rm flags when running your server

Solution 
version: "2"

agents:
  root:
    model: openai/gpt-4o
    instruction: Count the number of 'r' in a word
    toolsets:
      - type: mcp
        command: docker
        args: [run, "-i", "--rm", "mcp-strawberry"] # or djordjelukic1639080/mcp-strawberry

Our developer agent

Let's use the power of MCP servers to make our agent even more powerful.

A good developer always reads the documentation (right?, right?). So let's give our developer the capability to fetch up-to-date documentation for pretty much any library out there. For this we will use the context7 MCP server. Luckily for us, this server is already present in Dockers' MCP Toolkit.

version: "2"

agents:
  root:
    model: openai/gpt-4o
    instruction: You are an amazing developer
    add_environment_info: true
    add_date: true
    toolsets:
      - type: todo
      - type: shell
      - type: filesystem
+      - type: mcp
+        ref: docker:context7

Run this agent and ask this:

Create a directory "server" and in there write a server in node and typescirpt, the server should be a key-value store, use context7 to get the documentation you need for the server library you will use

Running this, the agent should use the context7 tools to find the appropriate documentation for the library it chose to use to write this server.

Look around the MCP Toolkit, play with some more MCP servers, what other servers do you think could be useful for a developer agent?