DevSpaces + Continue AI journey

In the browser IDE, Continue is the AI client. It talks to this chart in two separate paths: chat completions (LiteLLM) and optional Hub catalog tools (MCP Actions). Cluster Kubernetes tools stay on Hub — not in DevSpaces.

Mental model: AI chat ≠ MCP tools

Continue does not send chat messages through MCP. Chat and tools use different endpoints and secrets:

Che Code + Continue (DevSpaces workspace)
 │
 ├── Path A — AI chat / edit (Steps 06–07)
 │     Continue  ──litellm-master-key──►  LiteLLM Route /v1
 │                                         └──► Granite or Qwen (shared models)
 │
 └── Path B — Hub MCP tools (Step 08, optional)
       Continue  ──mcp-token──►  Hub Route /api/mcp-actions/v1
                                   └──► catalog / TechDocs tools only
  Path A — AI Path B — Hub MCP
What you prove Model replies in the Continue sidebar Tool call lists catalog Components
Endpoint LiteLLM Route /v1 Hub Route /api/mcp-actions/v1
Auth Secret rhdh-agent-sandbox-continue (LITELLM_API_*) mcp-token from rhdh-agent-sandbox-secrets
Wired by wire-continue → models: in ~/.continue/config.yaml Same command → mcpServers: hub-mcp-actions when Hub Route + token exist
Not available Native function-calling on shared Granite/Qwen OpenShift/Kubernetes MCP (ClusterIP only — Step 09)

How it feels in the IDE: you type in Continue. Path A always answers as a model. Path B only runs when you ask for catalog/TechDocs and Continue invokes an MCP tool (you should see a tool name such as query-catalog-entities, then a reaction with entity names).

What this journey proves

Path Endpoint Auth From DevSpaces?
A Continue chat → reply LiteLLM Route /v1 rhdh-agent-sandbox-continue Yes (Step 07)
B Continue MCP → catalog Hub Route /api/mcp-actions/v1 mcp-token Yes (Step 08)
Cluster OpenShift / K8s MCP ClusterIP services Hub Lightspeed No — Hub MCP Chat only

Shared Sandbox models via LiteLLM do not do native function-calling for Continue chat. Path B still works because Continue talks to Hub MCP Actions as an MCP client (tools are on Hub). Full K8s tool rounds stay on Hub with litemaas-qwen, or OpenClaw + LiteMaaS.

Wire Continue (v2): both paths in one command

Devfile command wire-continue loads secrets and writes:

  1. ~/.continue/config.yaml — Continue 2.x live config (models + optional mcpServers)
  2. Workspace .continue/config.json — placeholders only (safe to commit; no live keys)

Example Path B block (token from Secret, not from git):

mcpServers:
  - name: hub-mcp-actions
    type: sse
    url: https://<hub-route>/api/mcp-actions/v1
    requestOptions:
      headers:
        Authorization: Bearer <mcp-token>

oc get secret rhdh-agent-sandbox-secrets -o jsonpath='{.data.mcp-token}' | base64 -d

Try the same prompts

Path A — AI chat (Step 07):

Reply with exactly: DevSpaces Continue OK

Path B — Hub MCP (Step 08):

Use Hub MCP to list sample agent Components tagged agent.

Expected reaction names: sample-python-agent, sample-nodejs-agent, sample-quarkus-agent.

Stop when done

oc patch dw rhdh-agent-ai-demo --type=merge -p '{"spec":{"started":false}}'

Or Stop in the Dev Spaces UI.