Watch the walkthrough above without a cluster, or follow the live script below.

Interactive recording: Building Agentic AI in a Developer Sandbox (Arcade).

Use LiteMaaS Qwen (litemaas-qwen) for any tool-calling step. Shared Granite is chat-only — Lightspeed/Continue will hallucinate CVEs if you leave it selected in Agent mode.

Part A — Helm catalog install

1. OpenShift Console → Developer perspective → your Sandbox project
2. +Add → Helm Chart → search rhdh-agent-sandbox → Install Helm Chart
3. Release name rhdh-agent, chart 0.1.11, Form view
4. Set Cluster Router Base and OpenShift token (shared models) (oc whoami -t). Leave LiteMaaS API key empty if Secret litemaas-credentials already exists
5. Create → wait until Topology group rhdh-agent is Ready (Hub 2/2, LiteLLM, MCP, applier, PostgreSQL)

Part B — Hub agent loop (~5 min)

1. Open the Developer Hub Route → Enter as Guest
2. Catalog → filter mcp → browse MCP API entities and their Documentation
3. Create → Deploy Agent (Golden Path) (Python / LangGraph, tool-agent, model granite)
4. Open the new Component → Topology after the BuildConfig finishes (no git push)
5. MCP Chat → litemaas-qwen → ask for pods_list_in_namespace in this namespace

Part C — DevSpaces + Red Hat Security MCP (~5 min)

1. Create → Agent-friendly DevSpaces AI Workspace (workspace starts stopped)
2. Dev Spaces dashboard as the Sandbox user → Start → Trust
3. Continue → model LiteMaaS Qwen (LiteLLM) → Agent or Plan (not Chat)
4. Customer Portal SSO via the public http-8080 helper (not Che code-redirect)
5. Ask get_cve_by_id for CVE-2024-3094 (xz). Expect live tool chips, Red Hat severity — not NVD folklore
6. Stop the workspace when done to free quota

Wrap-up

oc get bc,deploy -l app.kubernetes.io/part-of=rhdh-agent-sandbox
oc get dw