Skip to content

Deploying Claude Code

Claude Code is Anthropic's agentic coding tool. The claude-code runtime is installed by the language-operator-runtimes chart. It exposes Claude Code's interactive CLI as a WebSocket terminal (xterm.js in the browser, node-pty on the server) so you can connect directly from any browser.

Claude Code can authenticate interactively (run /login in the terminal) or headlessly via an API key or subscription token injected from a Secret — see Authentication below.

Prerequisites

  • Language Operator installed, including the language-operator-runtimes chart (provides the claude-code runtime)
  • A LanguageCluster to deploy into, with your kubectl context set to its namespace (examples below assume a cluster named demo-cluster)
  • A Claude account (Pro, Max, Team, or Enterprise)

Instructions

Deploy Claude Code

kubectl apply -f - <<EOF
apiVersion: langop.io/v1alpha1
kind: LanguageAgent
metadata:
  name: code-agent
spec:
  runtime: claude-code
  instructions: |
    You are an expert software engineer. Help with code review,
    debugging, and implementation tasks.
EOF

Verify

kubectl get languageagents
kubectl get pods -w

Wait for the pod to reach Running and the LanguageAgent to show Ready=True.

Connect to the Terminal

The agent exposes a WebSocket terminal on port 8080. Port-forward to access it locally:

kubectl port-forward svc/code-agent 8080:8080

Then open http://localhost:8080 in your browser. You'll see Claude Code's interactive CLI running inside the pod.

To copy text, drag to select (hold Shift while dragging if Claude is in a mouse-tracking prompt), then Ctrl+C (smart copy: copies when there is a selection, otherwise passes through as SIGINT). Ctrl+Insert also works. On Mac: Cmd+C. Paste with Ctrl+Shift+V, Shift+Insert, or Cmd+V on Mac.

The terminal session persists across page reloads and pod restarts, and is shared across multiple browser tabs to the same agent — close your laptop, reopen, and you're back where you left off. The session is cleared only when Claude exits (e.g. /exit).

Authentication

Claude Code needs credentials to reach Anthropic's API. Pick one of the methods below. The non-interactive options inject a Secret value as an environment variable on spec.deployment.env, so agents come up authenticated with no per-pod setup — the right choice for headless or multi-agent deployments.

Interactive login (default)

If you set no credential env vars, open the agent terminal (see Connect to the Terminal) and run /login, completing the Claude.ai browser flow. Credentials are written to /workspace/.claude/.credentials.json and persist on the workspace PVC, so subsequent pod restarts don't re-prompt.

API key (headless)

Use ANTHROPIC_API_KEY for API-key billing — no /login needed:

kubectl create secret generic anthropic-credentials --from-literal=api-key=sk-ant-...

spec:
  runtime: claude-code
  deployment:
    env:
      - name: ANTHROPIC_API_KEY
        valueFrom:
          secretKeyRef:
            name: anthropic-credentials
            key: api-key

Subscription token (headless)

For subscription billing without a browser, mint a long-lived token with claude setup-token on your own machine, then inject it as CLAUDE_CODE_OAUTH_TOKEN:

claude setup-token   # prints a long-lived token
kubectl create secret generic claude-code-oauth --from-literal=token=<token>

spec:
  runtime: claude-code
  deployment:
    env:
      - name: CLAUDE_CODE_OAUTH_TOKEN
        valueFrom:
          secretKeyRef:
            name: claude-code-oauth
            key: token

See examples/development-team for a multi-agent setup that wires these credentials in (with optional: true so the same manifest works whether or not you provide them).

Configuration Reference

The claude-code runtime is configured through plain environment variables on the agent's spec.deployment.env. For example, to cap the number of agentic turns per request, set CLAUDE_CODE_MAX_TURNS:

spec:
  runtime: claude-code
  deployment:
    env:
      - name: CLAUDE_CODE_MAX_TURNS
        value: "10"

What the Operator Created

Resource Name Purpose
Deployment code-agent Runs the Claude Code WebSocket terminal container
Service code-agent ClusterIP on port 8080 (WebSocket terminal)
NetworkPolicy code-agent Allows inbound from other agents in this namespace. Add spec.networkPolicies.egress with cidr: 0.0.0.0/0 on port 443 to reach claude.ai, api.anthropic.com, and other public APIs.
PVC code-agent-workspace 10Gi persistent workspace at /workspace (also holds Claude config)
ConfigMap code-agent-agent Injected at /etc/agent/config.yaml