OpenStack MCP + Agent PoC

OpenStack MCP + Agent PoC

This repository contains the code and instructions to run an OpenStack MCP server and a very basic agent companion program to test it.

The approach taken in this PoC is:

• Use the OpenStack codegenerator project (openstack-code-generator) to create OpenAPI Specs for OpenStack.

• Use the mcp-openapi project to start multiple MCP servers (one per OpenStack component) from a filtered version of their full OpenAPI specifications.

• Use the agent code to interact with the MCP servers via an LLM.

You don't need all this for a quick test, you can just follow the QuickStart guide, but if you want to do something different you'll need to go through the Running things guide.

QuickStart

We assume you have access to an OpenStack deployment and to an LLM, as they are required to run things.

You'll also need the uv package manager. If you don't you can install it with sudo dnf install uv.

We'll be using 2 terminals, one for the MCP server and another for the Agent.

Clone this repository

We'll clone this repository with just the minimum external projects necessary to run things.

$ git clone --shallow-submodules \
  --recurse-submodules=mcp-openapi \
  https://github.com/Akrog/mcp-openstack.git

$ cd mcp-openstack

Run the MCP server

First we need to configure the MCP server, and for that we'll edit the servers.yaml file and replace the <NOVA_PUBLIC_URL>, <CINDER_PUBLIC_URL>, and <GLANCE_PUBLIC_URL> placeholders with the values of our cluster.

You'll notice that only a subset of REST API paths are being enabled to avoid having tool many tools that could go over our model's context window or induce hallucinations.

Now we can run the server:

$ cd mcp-openapi

$ uv run main.py --config ../servers.yaml

We leave this service running on this terminal and go to another terminal.

Run the Agent

In another terminal we'll go into the agent directory:

$ cd agent

OpenStack client config

Now we'll make sure the configuration for the OpenStack client is available in one of the standard locations as described in the documentation:

• agent/
• ~/.config/openstack/

If we are running an OSP 18 cloud we can get this with:

$ oc cp openstackclient:/home/cloud-admin/.config/openstack/clouds.yaml ./clouds.yaml
$ oc cp openstackclient:/home/cloud-admin/.config/openstack/secure.yaml ./secure.yaml 

LLM config

We'll create an llm.token file with the secret/token for the OpenAI LLM endpoint we want to use.

Then edit the agent.json file to replace the <LLM_URL> placeholder with our LLM's address and replace <MODEL> with the model we want to use. For example for Anthropic we could have something like this:

{
  "base_url": "https://api.anthropic.com/v1",
  "model": "claude-3-5-haiku-20241022",
  < ... >
}

Run the agent

Now we just need to start the agent:

$ uv run main.py

Enjoy

We can now ask things about our deployment and see the REST API calls on the MCP server terminal.

For example to get the nova flavors we could do:

$ uv run main.py
LLM: claude-3-5-haiku-20241022 @ https://api.anthropic.com/v1 Tools: 46
>>> What are my flavors?

Be aware that the agent doesn't have memory, so each prompt will be a clean prompt for the LLM.