Tuluat — Kubernetes-Native AI Operator
A Kubernetes-native AI operator and orchestration runtime that treats AI agents, LLM providers, and cognitive workflows as first-class cluster resources managed via GitOps.
Read the in-depth architectural deep-dive in the blog: Tuluat: Building a Kubernetes AI Operator in Java.
Declarative AI Infrastructure
Tuluat introduces declarative Custom Resource Definitions (CRDs) into Kubernetes:
LlmProvider: Centralized credentials, rate limits, and fallback configurations for foundational models.AiAgent: Scoped prompts, capabilities, persona definitions, and tool allowances.AiWorkflow: Composable multi-step or multi-agent graphs with conditional branching.WorkflowSession: State-persisted runtime instances of executing workflows.McpServer: Declarative endpoints for Model Context Protocol integrations.
Key Capabilities
- Durable Execution: Long-running multi-agent sessions survive crashes and support human-in-the-loop approvals (
WAITING_APPROVALstates) via the Temporal workflow engine. - Runtime Safety Guardrails: Automated PII redaction, prompt injection sanitization, and output confidence scoring built directly into the reconciliation layer.
- Tool Integration: Model Context Protocol (MCP) support over SSE and stdio transport registries.
- Vector Retrieval: pgvector retrieval with HNSW indexing over chunked documents stored in S3-compatible object storage.
- Turnkey Packaging: Packaged as a unified Helm chart including the operator, CRDs, Temporal, MinIO, PostgreSQL, Prometheus, and Grafana.
Technology Stack
- Core: Java 25 LTS (Virtual Threads), Spring Boot 4.1 (Spring Framework 7), Spring AI
- Operator & Orchestration: Java Operator SDK 5.1 (JOSDK), Temporal SDK, Embabel Goal Engine
- Storage & Vector: PostgreSQL + pgvector (HNSW), MinIO S3
- Tools & Protocols: Model Context Protocol (MCP SSE/stdio), WebSocket/STOMP, Helm