Python datasets, React and MCP

The Python client workflow serves ClickHouse datasets and fixed-measure metrics, generates TypeScript types/routes for the existing React hooks, and connects the Node MCP server to the same authorized HTTP API. Python owns SQL, authentication, tenant resolution and current endpoint policy.

This workflow is under development. Python remains pre-alpha; matching npm client-contract changes must be released together. Follow the runnable installed-artifact tutorial for pure definitions, offline export, authenticated serving, decimal/null seed, all four React query hooks, per-login cache isolation and MCP stdio configuration. The preparation command builds development artifacts outside the checkout. Published-version qualification remains a release gate.

The Python CLI exports build-time capabilities. The Node CLI's generate:client consumes that JSON and writes standalone ClientApi and route manifest types with endpoint schema pins. Build exports do not grant permission. Python reauthorizes every request; a schema change returns 409 and requires an explicit rebuild/reconnect.

The MCP CLI uses --client-contract-url for authorized Python HTTP discovery. The existing --self-test --url checks a hosted MCP endpoint. Host credentials belong in the environment, and one MCP process represents one backend principal. Browser code uses user-scoped credentials and resets cache scope on login changes.

The Python API needs no Node process. Frontend generation and MCP bridging require Node. Native Python MCP transport and hosted MCP OAuth are separate from this self-hosted workflow.