OpenStack LightSpeed MCPs
Exposes the full openstack and oc CLIs as MCP tools with read-only by default, dynamic/static credentials, and DNS rebinding protection.
README
OpenStack LightSpeed MCPs
This repository contains MCP tools that have been developed expressly for the OpenStack Lightspeed agent to use on an OpenStack on OpenShift deployment, so they are opinionated and functionality is tailored to our needs.
Current tools are:
openstack-cli: A simple tool to runopenstackCLI commands.openshift-cli: A simple tool to runocCLI commands. Requires theoctool to be installed.
Features
- Exposes the full
openstackandocCLIs as MCP tools - Read-only mode by default (configurable allow/block lists per tool)
- Dynamic credentials via HTTP headers — no credentials stored on disk required
- Static credentials from standard filesystem locations (
clouds.yaml,kubeconfig) - Bearer token authentication for the MCP transport layer
- DNS rebinding protection (mitigates CVE-2025-49596)
- Multi-worker support via stateless HTTP
- Configurable process pool for concurrent command execution
- Container image available at
quay.io/openstack-lightspeed/rhos-mcps
Why CLI tools?
Most MCP server implementations out there expose very narrow focused MCP tools following the precept that it is best to have each tool do a single thing, so if that's the recommendation, why are we doing a single MCP tool for the full CLI?
First of all, exposing the CLI in a tool is still a "single thing", even if the tool allows doing many things, it is still implementing just one functionality: Implement the openstack or oc command.
Even if you think that it is a technicality, which we don't, you must also think about this:
-
The
openstackclient has over 1570 commands, making it a herculean effort to implement as tools and the sheer number will make it unmanageable by the LLM. -
The Red Hat OpenStack documentation uses the
openstackcommands in its procedures, and having individual tools adds an abstraction layer, forcing the LLM to make the connection between theopenstackcommand that appears in the documentation and the tool. In contrast with our approach the LLM can just send the command from the documentation to the MCP tool.
Endpoints
Instead of having a single endpoint for the MCP server we have 2 different routes supporting Streamable HTTP MCP:
http://<server>/openstack/http://<server>/openshift/
The reason for this is to limit the exposure of the tokens, so each endpoint will only get the one it needs instead of both.
Credentials
The openstack-cli and openshift-cli tools need to know the endpoint to contact the cluster as well as credentials to access it just like the underlying CLIs they are exposing.
They support 2 mechanism to get the credentials:
- Static: The information is present on the filesystem in default locations.
- Dynamic: The information is passed as HTTP headers when calling the MCP tool.
Static
For openstack-cli the locations must be one of the supported ones by openstack:
./clouds.yamland./secure.yaml$HOME/.config/openstack/clouds.yamland$HOME/config/openstack/secure.yaml/etc/openstack/clouds.yamland/etc/openstack/secure.yaml- Location indicated by
OS_CLIENT_CONFIG_FILEenvironmental variable
For openshift-cli the locations must on the directory supported by oc:
$HOME/.kube/config- Location indicated by
KUBECONFIGenvironmental variable
Dynamic
For openstack-cli the headers are:
OS_TOKENOS_URL
For openshift-cli the headers are:
OCP_TOKENOCP_URL
Configuration
The configuration file is currently hardcoded to be ./config.yaml (a different location can be defined using RHOS_MCPS_CONFIG environmental variable) and if you want to see a valid configuration file the repository includes a sample configuration.
The configuration file has 4 sections:
- General
- OpenStack
- OpenShift
- MCP Security
General
ip: IP address the server will bind to. Default0.0.0.0.port: TCP port the server will bind to. Default8080.debug: Defaultfalse.workers: Number of different uvicorn workers. Default1.processes_pool_size: Maximum number of processes each worker can launch. Default10.log_format: Format string for the MCP server. Must use YAML's escape syntax, so\x1binstead of\033. Default:%(asctime)s.%(msecs)03d %(process)d \x1b[32m%(levelname)s:\x1b[0m [%(request_id)s|%(client_id)s] %(name)s %(message)suvicorn_log_format: Format string foruvicorn. Must use YAML's escape syntax, so\x1binstead of\033. Default:%(asctime)s.%(msecs)03d %(process)d \x1b[32m%(levelname)s:\x1b[0m [-|-] %(name)s %(message)s
MCP Security
These go under the mcp_transport_security key:
token: Token to use for basic authentication. The caller must send this value in request headerAuthorization. For example:Authorization: Bearer supersecret. Default"".enable_dns_rebinding_protection: Enable DNS rebinding protection")allowed_hosts: List of allowed hosts. Default["*:*"]allowed_origins: list of allowed origins. Default["http://*:*"].
With the exception of the token field the rest comes from the Anthropic MCP SDK.
For example the DNS rebinding protection was introduced by Anthropic to mitigate critical RCE vulnerabilities (e.g., CVE-2025-49596). The protection restricts unauthorized local service access by verifying Origin and Host headers, requiring session tokens, and blocking malicious DNS-to-localhost mappings.
OpenStack
These go under the openstack key:
allow_write: Whether to allow write operations or not. Defaultfalse.ca_cert: CA certificate bundle file location. Default"".insecure: Whether to allow insecure SSL connections or not. Defaultfalse.
OpenShift
These go under the openshift key:
allow_write: Whether to allow write operations or not. Defaultfalse.insecure: Whether to allow insecure SSL connections or not. Defaultfalse.allowed_commands: List of allowed commands whenallow_write: false. Example:["status", "projects", "get", "adm top"].blocked_commands: Explicitly blocked commands whenallow_write: true. Example:["get-token", "logout"].
Running the service
We assume the configuration file has already been created, for example config.yaml before we run the server using one of the following methods:
From source code
pdm run rhos-ls-mcps
Container Image
The container image is available at quay.io/openstack-lightspeed/rhos-mcps.
We assume for the following examples that you haven't set the port in our config file or we've set it to the default 8080.
Dynamic credentials:
podman run -p 8080:8080 \
-v ./config.yaml:/app/config.yaml:Z \
quay.io/openstack-lightspeed/rhos-mcps:latest
Static credentials:
We assume we are using CRC for the deployment, so we need to get the openstack credentials from the OpenShift cluster:
scripts/get-crc-creds.sh
This creates 3 files in our current directory: clouds.yaml, secure.yaml, tls-ca-bundle.pem.
And we have the OpenShift configuration and credentials in ~/.crc/machines/crc/kubeconfig.
So we can run the service using all these files and our config.yaml file where we define the location of our ca-bundle.
podman run -p 8080:8080 \
--user 1001 \
-v ./config.yaml:/app/config.yaml:Z \
-v ./clouds.yaml:/app/clouds.yaml:Z \
-v ./secure.yaml:/app/secure.yaml:Z \
-v ./tls-ca-bundle.pem:/app/tls-ca-bundle.pem:Z \
-v ~/.crc/machines/crc/kubeconfig:/opt/app-root/src/.kube/config:Z,U \
quay.io/openstack-lightspeed/rhos-mcps:latest
Checking the servers
Assuming port is 8080 and the secret token used for the server is supersecret we can run a single curl command for each endpoint to confirm it's running and responds to requests:
curl -X POST http://127.0.0.1:8080/openstack/ \
-H "Content-Type: application/json" \
-H "Accept: application/json, text/event-stream" \
-H "Authorization: Bearer supersecret" \
-d '{"jsonrpc":"2.0","id":"1","method":"tools/list"}'
curl -X POST http://127.0.0.1:8080/openshift/ \
-H "Content-Type: application/json" \
-H "Accept: application/json, text/event-stream" \
-H "Authorization: Bearer supersecret" \
-d '{"jsonrpc":"2.0","id":"1","method":"tools/list"}'
Additionally we can use the MCP inspector tool against the server to run commands.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.