A SSPCloud compute workstation, driven by your AI assistant.
SSPCloud MCP gives Claude (Desktop, mobile, Code) a remote workstation on the datalab: push a repo, run code, use the GPU, stateful notebooks, and deploy services — all from the conversation, in your namespace.
The problem
An AI assistant can write code, but not run it on a real machine. Whenever a GPU, a notebook or a heavy job is needed, the conversation stops at "launch it yourself from the interface". Yet the SSPCloud datalab offers free compute — the bridge was missing.
What it does
Two verbs, remotely, on the datalab's real infrastructure.
Run code
Stateful Python kernel (variables, GPU model in memory) + shell: git, pip, pytest in the pod.
Arbitrated GPU
One GPU slot per user, switched per project with preemption; each project's environment preserved (PVC).
Full notebooks
Create, run, iterate a .ipynb; your colleague reopens the same file in their JupyterLab.
Deploy a service
Expose an API or a demo over HTTPS from the conversation. 24 MCP tools in total.
Durable & secure
A "product" service, not a fragile hack that falls over.
Durable
The server is the main process of a Kubernetes Deployment: auto-restart, never auto-suspended, survives closing your pods.
Browser OAuth
OAuth 2.1 + DCR + PKCE MCP connector (with CORS) for Claude Desktop, mobile and claude.ai. Or a static header for Claude Code / Cursor.
Single-namespace
Everyone installs their service in their namespace. Auto-generated bearer, never known to the author. No access to others' data.
Install
Two paths — the simplest is one click from the Onyxia catalog.
kubernetes.role: edit, one command (above) creates the durable Deployment and prints your key.…/mcp URL into a connector; at the OAuth form, enter the key. The 24 tools appear.