“Hey AI, Train Llama”: Making Kubeflow Agent-Native With MCP
Talkat Open Source Summit India 2026 · on Wednesday, 2026-06-17· at 12:50· Lotus Ballroom 3 (Level 3)·
summary
Kubeflow MCP Server exposes the Kubeflow SDK as MCP-callable tools so agents can submit and manage training via natural language while enforcing cluster safety. Prototype maps Kubeflow operations into 20 tools and 3 resources, runs demos with LangChain + Ollama and a remote Qwen 27B, integrates MLflow and Jaeger/OpenTelemetry for traces, performs pre-flight checks, enforces persona-based filtering, and requires a two-phase confirmation before large GPU allocations; project is alpha with pre-release work and CNCF onboarding.
AI-written summary · about
chapters (11)
- 0:00Kubeflow agent-native intro
- 5:41Kubeflow ecosystem overview
- 9:15Kubeflow MCP for AI agents
- 11:46Security in MCP workflows
- 12:50Demo: Light LLM agent run
- 16:03Tool filtering by persona
- 17:26Debugging training workflows
- 22:50Demo: Parallel runs Cloud/Cursor
- 31:05Roadmap for Kubeflow MCP
- 33:52Call for contributors
- 34:54Q&A
related talks
44mEngaging the Kubeflow Community: Building an Enterprise-Ready AI/ML Platform
16mStreamline LLM Fine-tuning on Kubernetes With Kubeflow LLM Trainer
32mKubeflow Ecosystem: What’s Next for Cloud Native AI/ML and LLMOps
21mStreamline LLM Fine-Tuning on Kubernetes with Kubeflow LLM Trainer
31mKubeflow 2026 State of the Union: Orchestrating the Full ML Lifecycle
21mIntroducing Kubeflow (w. Special Guests Tensorflow and Apache Spark)
14mKubeflow: Where ML Meets Kubernetes
15mAgentic RAG on Kubeflow