[kubeflow] [ANNOUNCE] Kubeflow MCP: talk to your ML platform like it's an agent

[Abhijeet Dhumal]

Hi Kubeflow Community,

We’re excited to introduce the Kubeflow MCP 

What started as a prototype 4 months ago is now a CNCF-adjacent community project backed by KEP-936. — and was just selected for Google Summer of Code 2026 (200+ applicants, ranked 2nd based on  positive traction).

git clone https://github.com/kubeflow/mcp-server.git
cd mcp-server && pip install .

What it does: natural language -> Kubeflow — no YAML, no SDK boilerplate. Works out-of-the-box with Cursor, Claude Code, and any MCP-compatible agent.

Model registry, Pipelines, Feast, Spark Connect and KServe are next.

That said - this is still early. The foundation is solid, but the real potential depends on how the community shapes it.
If you can file an issue, review a PR, test it against a real cluster, or just share it with someone who might care, that's exactly the kind of support that moves this in the right direction. Every bit counts.

If you haven't already - give us a   on GitHub and share it with anyone training using agentic workflows.

What's next : Kubeflow MCP 2026 Roadmap
 GitHub: https://github.com/kubeflow/mcp-server

For additional questions, please join the #kubeflow-ml-experience channel.

Regards,
Abhijeet Dhumal