The problem
Data Cloud spreads its configuration over many screens and APIs. I wanted to ask an assistant "which streams feed this data model object?" or "propose a mapping for these fields" and have it work against the real org.
What it does
- More than 30 tools covering data streams, segments, data model and data lake objects, transforms, calculated insights, identity resolution, data graphs, activations, connections, data spaces and the Profile and Query APIs.
- AI-assisted write tools: propose and then apply field mappings, or propose the SQL for a calculated insight and create it.
- Three ways to sign in: client credentials, JWT bearer, or an OAuth browser login.
- Local or remote: run it with
npx -y salesforce-datacloud-mcp, or deploy the remote HTTP mode to Cloudflare Workers.
How it is built
Data Cloud authentication is a two-step exchange: a Salesforce OAuth token first, then a Data Cloud token swapped for it. The server handles both and refreshes tokens before they expire. In remote mode, credentials travel as request headers on each call so they stay on your machine and are not stored on the server.
What I learned
- Plan less, ship the tools people use. The first planning document listed more tools than I built. The ones that survived are the ones I reached for in real work.
- Propose, then apply. Splitting write operations into a proposal step and an apply step keeps a human in the loop for anything that changes the org.