Karim Boumedhel karmab

Connecting a Coding Agent to OpenShift AI: What Actually Works (and What Doesn't)

Inspired by Frédéric Klein's article on connecting Claude Code to a DGX Spark via Red Hat AI.

TL;DR

We connected OpenCode (an open-source coding agent CLI) to models served by Red Hat OpenShift AI (RHOAI) on an OpenShift cluster with NVIDIA L4 GPUs. We got end-to-end tool calling working — the agent can run bash commands, read files, and edit code. But the honest answer is: it works for a demo, not for daily use. Here's everything we learned so you don't have to repeat our mistakes.

Background: Where Frédéric's Article Left Off

a plan to create:

a hub openshift cluster (on a sno)
an sno spoke vm
a workflow deploying AI operator

Steps to create a virtual cluster with baremetal workers

requirements

valid pull secret
a bridge created on top of your default interface in the hypervisor, in this documentation, the bridge is named baremetal

Create base cluster (using a param file like the one provided as an example)

A plan to deploy openshift on a dedicated (libvirt) ipv6 network

How to use

kcli create plan -f kcli_plan_ipv6.yml

or, if overriding some of the parameters

this is a simple tool called federer.py which, for a given yaml file:

splits its multiple documents
create the corresponding federated and placement files
indicate which types will be required
a script to actually deploys

The tool has also won 20 grand slams

	package main

	import (
	"log"
	"os"

	"golang.org/x/net/context"
	"golang.org/x/oauth2/google"
	"google.golang.org/api/compute/v1"
	)

	import argparse
	import os
	import sys

	PARAMHELP = "specify parameter or keyword for rendering (multiple can be specified)"


	def get_subparser_print_help(parser, subcommand):
	subparsers_actions = [
	action for action in parser._actions

	apiVersion: tekton.dev/v1beta1
	kind: PipelineRun
	metadata:
	generateName: mysupertask
	spec:
	pipelineRef:
	name: mysupertask
	workspaces:
	- name: sourcedir
	volumeClaimTemplate:

	apiVersion: v1
	kind: ReplicationController
	metadata:
	labels:
	app: kubevirt-web-ui
	openshift.io/deployment-config.name: kubevirt-web-ui
	name: kubevirt-web-ui
	spec:
	replicas: 1
	selector: