2026 · Backend, infrastructure, MCP
Kaggle Cloud Manager
A web dashboard and job queue for running Kaggle notebook executions remotely, with the same control plane exposed to any AI agent over MCP.
- Python
- FastAPI
- Postgres
- MCP

A closer look
Kaggle Cloud Manager
Commits in the last year
17
Language bytes
- Python120k
- Kotlin48k
- JavaScript40k
- CSS14k
- HTML3k
The problem
A Kaggle notebook run is a browser tab you have to keep open, on a free account where the queue is the real cost. There is no queue of your own, no log you can keep, and no way to ask anything else to start one.
What I built
- 01
One job runs at a time, enforced in the database rather than in memory, so a redeploy cannot start a second execution.
- 02
Job state lives in Postgres, which means a queued run survives a restart and an interrupted run can be requeued.
- 03
Kaggle is driven through its REST API directly, with no heavyweight SDK and full control over retries.
- 04
The MCP server exposes the same tools the dashboard uses, so an agent can list notebooks, run one, and read the logs without a browser.
- 05
It is built for a free web service tier: one process, a hosted database, no build step and no on-instance storage.
What it changed
One execution at a time, guaranteed by the database.
Stop works by pushing an immediate-exit version of the notebook, and the outputs survive the stop.
The agent path and the dashboard path are the same API.