The workflow MCP removes
The normal shape of an analytical question is: form the question, leave your assistant, find the data, download the data, clean the data, load it, chart it, come back, write it up.
Most of that is transport. The analysis is the first step and the last one.
MCP collapses the middle. The assistant calls the data source directly, gets structured rows back, and reasons over them in the same conversation where you asked. You do not leave, and you do not hand-carry a CSV between two tools that both understand the data perfectly well.
What that changes in practice
It changes what questions are worth asking.
When pulling a dataset costs twenty minutes, you only pull it for questions you already know are important. When it costs one call, exploratory questions become cheap — and exploratory questions are where most of the actual findings live. "Which counties look unusual" is a bad use of twenty minutes and a good use of one call.
The honest problem: verification
The risk in this arrangement is obvious once stated. A model that receives no data and a model that receives empty data both produce fluent prose. If a tool silently returns nothing, you get a confident summary of nothing, and nothing in the output looks wrong.
This is not hypothetical — it is the same class of failure as a scraper that returns a rendered error page with a 200 status. The difference is that here, a language model sits downstream, and language models are extremely good at making thin input sound substantial.
Two things help:
Tools should say what they returned. A row count, a coverage statement, an "as of" date. If a tool hands back rows with no idea how many were expected, neither you nor the model can tell a genuine empty result from a broken one.
Free previews should be real. A masked preview that returns actual live rows with the identifying field hidden proves the pipeline works end to end. A preview that returns canned sample data proves nothing at all, and is worse than none because it is reassuring.
How ours is set up
Connect an MCP-capable client to `product.kcalbinmcp.uk/mcp` and the tools are available immediately, with no account:
- County displacement signals and WARN notices — severity, forecast and filings, with county and company names masked until you subscribe. - Death-care demand — expected annual deaths per county and year-by-year projections to 2040. - `browser_fetch` — renders a public page in a real headless browser and returns title, text and links.
The masking is the point. You see real rows from the live dataset, counted and dated, with the identifying column withheld. You can confirm the data exists and that the connection works before any money changes hands — and if the pipeline were broken, the preview would fail visibly rather than quietly.
What it is not
It is not a replacement for a warehouse. If you need to join these datasets against your own at scale, take a bulk extract and load it properly. MCP is for the question you have right now, in the place you are asking it.