Take-home · Kubernetes

In Lab 2 you wrote an MCPRoute that puts five MCP servers behind one endpoint, filters them down to 8 tools and authorizes every call per agent identity. Here you deploy that same route to a local Kubernetes cluster (kind) running Envoy Gateway and Agent Router.

The policy doesn't change. Only two references to infrastructure do: which Gateway the route attaches to, and how it finds the servers. What works on your laptop deploys unchanged.

Everything for this page is in takehome/kubernetes/ on GitHub.

Before you start

Run this on your laptop. The workshop Codespace has no Docker.

• Docker, running, with about 4 GB of memory free
• kind v0.24 or newer (v0.27+ recommended), kubectl, and helm v3
• Node.js 24 (or 22.18+) and the repository's dependencies (npm ci), for tokens and the sample agent

1. Bring it up

From the repository root:

./takehome/kubernetes/up.sh

The first run takes 2 to 5 minutes, mostly downloads. Re-runs take about 20 seconds. The script:

1. creates the kind cluster agent-router-workshop (Kubernetes 1.32),
2. installs Envoy Gateway v1.8.1 and Agent Router v1.1.0 with Helm, as in the Agent Router installation guide,
3. builds the MCP servers image and loads it into kind,
4. applies the manifests and waits until a request without a token gets a 401.

Then, in a second terminal, forward the router to your laptop and leave it running:

kubectl -n envoy-gateway-system port-forward svc/agent-router 8080:80

2. Check that it behaves like Lab 2

You don't need the local router (./lab run) for any of this.

No token: a 401, with a pointer to the OAuth metadata, exactly as in Lab 2:

curl -si http://localhost:8080/mcp -H 'content-type: application/json' \
  -H 'accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' | grep -iE '^HTTP|www-authenticate'

Then run the sample agent against the cluster:

TOKEN=$(./lab token release-bot)
node agent/index.ts --mcp http://localhost:8080/mcp --token "$TOKEN" --list              # 8 tools
node agent/index.ts --mcp http://localhost:8080/mcp --token "$TOKEN" --brain scripted    # deploys to staging
node agent/index.ts --mcp http://localhost:8080/mcp --token "$TOKEN" --brain scripted \
  --task "Deploy checkout 1.4.3 to production"                                           # denied (403)
node agent/index.ts --mcp http://localhost:8080/mcp --token "$(./lab token triage-bot)" --list   # 6 tools, no deploy__*

This cluster has no LLM route. To use an LLM instead of the scripted brain, point the agent at your provider directly, for example AGENT_LLM_BASE_URL=https://api.openai.com/v1 AGENT_LLM_API_KEY=sk-… AGENT_MODEL=gpt-4.1-mini. Any other MCP client works too: use http://localhost:8080/mcp with an Authorization: Bearer header, as on Bring your own agent.

3. Change the route

Edit takehome/kubernetes/manifests/mcproute.yaml and apply it. The router picks up the change in a few seconds, with no restart:

kubectl apply -f takehome/kubernetes/manifests/mcproute.yaml

Watch the access log, one JSON line per MCP call:

kubectl -n envoy-gateway-system logs deploy/agent-router -c envoy -f

What's different from the local labs

	Local labs (./lab run)	This take-home
Router	aigw run, one process on your laptop	Envoy Gateway and Agent Router controllers; an Envoy pod with the MCP proxy beside it
Endpoint	http://localhost:1975/mcp	Service agent-router:80, port-forwarded to http://localhost:8080/mcp
MCP servers	node servers/index.ts on ports 3001–3005	The same code in one Deployment, with Services issues, ci, deploy, docs and chat
Gateway	aigw-run	agent-router
MCPRoute namespace and parentRefs	default, aigw-run	lakeshore, agent-router
MCPRoute backendRefs	kind: Backend: an Envoy Gateway Backend at 127.0.0.1:300x	A Service name and its port (no kind means Service)
toolSelector, oauth, authorization rules		Unchanged
Applying a change	Edit workspace/mcproute.yaml, then ./lab run	kubectl apply -f manifests/mcproute.yaml

The only changes to the route itself, from the Kubernetes manifest:

metadata:
  name: ship-it
  namespace: lakeshore          # was: default
spec:
  parentRefs:
    - name: agent-router        # was: aigw-run
      kind: Gateway
      group: gateway.networking.k8s.io
  path: /mcp
  backendRefs:
    - name: issues
      port: 3001                # was: kind: Backend, group: gateway.envoyproxy.io
      path: /mcp
      toolSelector:
        include: [get_issue, add_comment]
    # … ci, docs, deploy, chat: the same change
  # securityPolicy: unchanged from Lab 2

Compare the full file with Lab 2's solution. Each changed line is marked CHANGED.

About the token audience

The route still expects aud: http://localhost:1975/mcp and advertises it as its resource, because that's what ./lab token puts in every token. That's also why the 401's resource_metadata link says :1975. In production, audiences and protectedResourceMetadata.resource are your endpoint's real public URL, say https://mcp.example.com/mcp. Your authorization server issues tokens for that URL, and remoteJWKS (or discovery from the issuer) replaces the workshop's JWKS ConfigMap.

Troubleshooting

• Port 8080 is taken. Forward another port (8081:80) and use it in the URLs.
• kind load says "failed to detect containerd snapshotter". kind older than v0.27 can't load images into newer node images. up.sh imports the image with ctr instead, so you can ignore it.
• Something isn't ready. Check kubectl -n lakeshore get pods,gateway,mcproute and the router's logs: kubectl -n envoy-gateway-system logs deploy/agent-router -c ai-gateway-extproc.

Clean up

./takehome/kubernetes/down.sh        # deletes the kind cluster
docker rmi lakeshore-servers:workshop