Baker Street

Overview

What is Baker Street and why does it exist?

Overview

Every AI assistant today works the same way: you open a chat window, type a message, get a response, and the conversation disappears. The assistant lives in a sandbox with no memory of who you are tomorrow, no ability to act on your infrastructure, and no way to work in the background while you sleep. It is a conversation partner, not an agent.

Baker Street flips this model. Instead of an AI trapped inside a chat box, Baker Street gives the AI an entire operating environment: Kubernetes. The agent can dispatch background jobs, remember things permanently, reach out through Telegram or Discord, launch isolated task pods for complex work, discover new tools at runtime, and coordinate with companion agents running on other machines.

The key insight: the application itself becomes the prompt. The personality files, the tool configuration, the memory system, the extension ecosystem -- all of it shapes how the agent behaves. Change the app, change the agent.

Who is Baker Street for?

Baker Street is built for developers, platform engineers, and homelabbers who want an AI agent that runs on infrastructure they control. If you are comfortable with Kubernetes, enjoy tinkering with infrastructure, and want an agent that grows with you rather than resetting every session, Baker Street is your platform.

Key Principles

Self-Hosted

Baker Street runs entirely in your own Kubernetes cluster. Your data, your models, your rules. There is no SaaS dependency beyond the LLM API itself. Conversations, memories, and job history stay on your infrastructure.

Kubernetes-Native

Every component runs as a Kubernetes pod, deployment, or job. Baker Street uses Kubernetes primitives for isolation, scaling, scheduling, and secret management. Network policies enforce default-deny ingress. Pod security contexts drop all Linux capabilities. This is not an app that happens to run on Kubernetes; it is an app that leverages Kubernetes as a core part of its architecture.

Security-First

Security is a first-class architectural concern. Every container runs as non-root with a read-only filesystem. Secrets are scoped so each service only sees what it needs. Task pods execute with zero RBAC permissions, no ingress, and automatic cleanup. The enterprise distribution adds guardrail middleware, HMAC-chain audit logging, and supply chain verification.

Extensible by Design

Baker Street has four tiers of extensibility. Write a Markdown file to teach the agent something new (Tier 0). Build a TypeScript plugin for in-process tools (Tier 1). Deploy a Kubernetes pod that speaks MCP to add capabilities at runtime (Tier 2). Run a companion daemon on bare metal to extend the agent's reach beyond the cluster (Tier 3).

Core Capabilities

  • Conversational Agent -- Claude-powered reasoning with multi-step tool dispatch and real-time streaming responses
  • Long-term Memory -- Qdrant vector store with Voyage AI embeddings for semantic recall across conversations
  • Background Jobs -- NATS-backed workers handle agent tasks, shell commands, and HTTP requests in parallel
  • Task Pods -- Ephemeral Kubernetes jobs with zero ingress for isolated code execution
  • Extensions -- Deploy a pod, gain a tool. MCP-protocol extensions are discovered automatically at runtime
  • Multi-channel -- Web UI, Telegram, and Discord adapters with per-channel conversation isolation
  • Observability -- OpenTelemetry traces, Prometheus metrics, and structured logging with Grafana dashboards

What's Next?

Read the Architecture guide to understand how the components fit together, or jump straight to the Quickstart to deploy Baker Street in your cluster.