Before You Begin¶
This tutorial deploys real services to a real cluster. Before you start Module 1, work through this checklist. Each item links to a setup guide if you need help.
Two paths through this tutorial
Path A — Full cluster. You have an OpenShift cluster with OpenShift AI
installed and at least one GPU available. You will serve gpt-oss-20b
on-cluster with vLLM and route the agent to it. This is the intended
experience.
Path B — External model. You don't have a cluster, or your cluster has
no GPUs (e.g., the Red Hat Developer Sandbox). You will still
deploy the agent, MCP server, gateway, and UI — but the LLM lives somewhere
else. Any OpenAI-compatible endpoint works (a hosted vLLM, a corporate
inference gateway, etc.). Wherever the tutorial says "set MODEL_ENDPOINT,"
point it at your external URL.
Requirements¶
1. An OpenShift cluster¶
OpenShift 4.20 or later. You need cluster-admin (or equivalent permission to install operators and create namespaces).
→ See Choosing a Cluster for self-managed, ROSA, Developer Sandbox, and local CRC tradeoffs.
FIPS mode is encouraged
Every agent and MCP server you build in this tutorial is FIPS-compatible. If you're standing up a fresh cluster, consider enabling FIPS at install time — it can't be turned on later. See the Red Hat OpenShift FIPS documentation for instructions.
2. Red Hat OpenShift AI and GPU support (Path A only)¶
The Red Hat OpenShift AI operator (3.2 or later) must be installed via
the fast-3.x channel, and a DataScienceCluster provisioned with KServe
enabled for model serving. On-cluster model serving also requires at least
one GPU-enabled worker node. RHOAI 3.x requires OpenShift 4.20+.
→ See Install OpenShift AI and GPU Support.
4. An LLM (RedHatAI/gpt-oss-20b)¶
The tutorial uses RedHatAI/gpt-oss-20b served via vLLM. You need:
- An OpenAI-compatible endpoint URL (
MODEL_ENDPOINT) - The model identifier (
MODEL_NAME)
Path A: deploy vLLM on your cluster (one ~24 GB GPU; the MXFP4-quantized variant fits in ~16 GB). Path B: use any external OpenAI-compatible URL.
→ See Serve an LLM. The Path B fallback is in the same guide.
5. CLI tools¶
oc(OpenShift client)helm3.xpipx- Python 3.11 or later
fips-agentsCLI (pipx install fips-agents-cli)
→ See Install CLI Tools.
6. A container registry you can push to¶
Modules 2, 3, and 6 build and push container images. Quay.io (free public namespaces) works for the tutorial; the OpenShift internal registry works if you'd rather keep everything in-cluster.
→ See Registry Setup.
Quick verification¶
Before starting Module 1, you should be able to run:
oc whoami # logged into the cluster
oc get dsc # DataScienceCluster exists
curl -s "$MODEL_ENDPOINT/v1/models" # LLM responds
fips-agents --version # CLI installed
helm version --short # helm available
If any of those fail, fix it before continuing — the tutorial assumes all five work.
Multi-cluster safety
Throughout the tutorial, every oc command includes --context="$CTX" to
avoid accidentally targeting the wrong cluster. Set it once per shell
session:
Ready?¶
Head to Module 1: Scaffold Your Agent.