Self-hosted AI agents.
From a prompt.

Describe what you want. It deploys to your cluster. Your keys, your data, your namespace.

terminal
$brew install tiny-systems/tap/mcp-server
🍺 /opt/homebrew/Cellar/mcp-server/0.1.6: built in 1 second
$claude
>
Install Now View on GitHub

What can the agent do?

Real prompts. Real agents. Running in your cluster, not ours.

A K8s incident triage agent that watches my pods and pings me on Slack when something looks wrong, with a one-line LLM diagnosis

Pod Health Agent
·pod_watch + llm_complete + slack_send

A support chatbot with memory that answers questions over my internal docs, no SaaS in the loop

RAG Chat Agent
·http_server + embed_text + llm_chat + document_store

A daily triage agent that reads my open GitHub issues and proposes labels at 9am

Issue Triage Agent
·cron + http_request + llm_tools

An on-call agent that takes a question on Slack, queries Prometheus, and explains what looks off

On-call Agent
·slack_command + llm_tools + http_request
Explore more agents

Why teams who run K8s pick this.

Production-shaped from day one. Built for the cluster you already operate.

⎈

Kubernetes-native

Real CRDs, real operators, real helm. Multi-replica, namespace-isolated. Built for the cluster you already run, not Docker Compose pretending to scale.

🤖

Built from a prompt

Describe the agent in Claude Code, Cursor, or any MCP-compatible editor. The MCP server composes the flow, picks modules, wires edges. No drag-and-drop tour.

📦

Capabilities, helm-installable

Every agent capability — LLM, vector store, K8s actions, HTTP, Slack — is a public Helm chart in the catalog. Anyone publishes, anyone installs. Bundle TEI or pgvector as subcharts on the same release.

🔒

Your keys, your data

Nothing leaves your namespace. API keys live in K8s Secrets. No SaaS proxy, no telemetry phone-home, no third party reading what your agent reads.

🧩

Multi-provider, no lock-in

Anthropic by default, OpenAI and any compatible endpoint (Ollama, vLLM, OpenRouter) via a single setting. Swap models per node, mix providers in one flow, keep your data behind your firewall.

⚡

Production-grade

Leader election, metadata persistence, graceful shutdown, exponential backoff, OTel traces. The runtime your platform team will actually sign off on.

Get started in 30 seconds.

Full setup guide — multi-cluster, JSON configs, all AI clients.

# Install the MCP server
$ brew install tiny-systems/tap/mcp-server
# Connect to Claude Code
$ claude mcp add tinysystems mcp-server serve
# Start building
$ claude

Also available via Go install or binary download. Works with any MCP-compatible client.

Your first agent in your cluster.

One Helm install. One prompt. The agent runs where your data already lives. No SaaS in the loop. No drag-drop tour.

Install Now Read the docs