Klaus

klaus

An [MCP](https://modelcontextprotocol.io/) server that wraps [Claude Code](https://docs.anthropic.com/en/docs/claude-code) to run AI coding agents inside Kubernetes.

Overview

Klaus runs the Claude Code CLI as a managed subprocess and exposes it over HTTP as a Streamable HTTP MCP endpoint. This allows orchestrating AI coding agents in Kubernetes with proper lifecycle management, health checks, and optional OAuth authentication.

MCP Client --> /mcp --> MCP Server --> Prompter --> Claude Code CLI (subprocess)

MCP Tools

Tool	Description
prompt	Send a task to the Claude agent. Non-blocking by default (set blocking=true to wait for completion).
status	Query agent state, progress, and result. Returns completed with result when a non-blocking task finishes.
stop	Terminate the running agent
result	Get full untruncated result and message history from the last run (debugging tool)

Extension system

Klaus supports the full Claude Code extension surface, configured via Helm values, klausctl config, or operator CRDs:

• Skills -- Domain knowledge loaded as SKILL.md files
• Subagents -- Specialized agents for task delegation
• Hooks -- Lifecycle hooks for validation and automation
• MCP Servers -- External tool integrations
• Plugins -- OCI-distributed extension bundles

Quick start

export ANTHROPIC_API_KEY=sk-ant-...
klaus serve
# MCP endpoint available at http://localhost:8080/mcp

Or with Docker:

docker run -d -p 8080:8080 -e ANTHROPIC_API_KEY=$ANTHROPIC_API_KEY gsoci.azurecr.io/giantswarm/klaus:latest serve

Or deploy to Kubernetes:

helm install klaus helm/klaus/ -n klaus \
  --set anthropicApiKey.secretName=anthropic-api-key

Documentation