YOUR ORGANIZATION'S MCP
Most people don't get what a governed MCP enables, so we built a game to explain it.
Walk up to anyone and ask them a question about the business. See how far that gets you.
ON THE STREET
Ask five people in a company a question about the business and you get five ways of not answering it. Here are the five, verbatim from the calls they came from.
-
THE ANALYST
Sure. I can pull that. I will get to it Thursday. There are nine ahead of you, and every one of them needs a chart before Monday.
-
DATA ENG
We can build that. It takes forever, and then nobody understands what we are talking about, because you are on rev ops and we are engineers.
-
THE VENDOR
Great question. Our platform absolutely does that. It does everything, honestly. It kind of sucks at everything equally, but it does everything.
-
DASHBOARDS
Which one do you want? There is one for 90 days, one for 180, and one for the year. They disagree. That is sort of the whole problem with dashboards.
-
THE CHATBOT
Here you go. An incredible report. It is an HTML file. Go nuts. It has nowhere to live, no way to refresh, and no way to share.
THE DATA OS
Same four questions, answered from the governed tables with the definitions applied first.
-
What was the conversion rate of Qualified Lead to sale vs non qualified lead?
Before the number: your glossary defines a Lead as someone who submitted the quote request with name, address and contact details. Qualified is your CRM's disposition field. It is not a score I invented on the way to answering you.
Qualified leads converted to a signed job at 30.0 percent. Unqualified converted at 3.9 percent. Same period, same definition, 3,458 leads between them.
That is a 7.7x gap sitting underneath a blended rate of 13.0 percent. The blend looks stable month to month because the mix moves and the average hides it.
- Used your definition of Lead
- Read the governed tables
ARTIFACT: Qualified vs unqualified, by month. Table and chart, with the definition it applied printed on the front.
-
How are we pacing forecast to target for rev and margin?
Revenue: you are at 71 percent of this month's target with 9 days left, which puts you on pace for 96 percent. The target is not my guess. It was set in your glossary in January and it has been sitting there since.
Margin: I cannot answer that one. Your cost of goods is not in the warehouse. Nothing upstream carries it, so there is no honest way for me to produce the number.
I could estimate it. I am not going to, because you would put the estimate in a board deck and by Thursday it would be the number. Connect a cost source and this becomes one query.
- Target read from your glossary
- Named what it does not have
PACING: Revenue pace to target. Revenue only. The margin gap is written on the front, not buried in a footnote.
-
What happened yesterday?
Yesterday meaning the 24 hours ending at midnight in your reporting timezone. I say the range out loud every time, so a wrong assumption is visible instead of buried.
Spend, sessions, quote starts, leads and signed jobs all landed inside their normal band. One did not. Quote starts came in at 17.4 percent against the 18 percent floor you set.
That is day one of three, and your alert fires on the third. I can put it in Slack now instead of waiting, if you would rather know early.
- Range named in the answer
- Reads the governed tables
DIGEST: Yesterday, on one page. Five metrics, one flag, and the Slack alert already written for you.
-
Why did quality drop?
Quality did drop, and it dropped in the way that is easiest to miss: lead volume held while the rate of leads becoming jobs fell. On a volume report this month looks fine.
The mix moved. Meta Ads sent 878 leads that became 47 signed jobs. Email sent about a fifth of that traffic and produced 94. Your team did not get worse at selling. You bought more of the cheaper lead.
One thing I cannot tell you: which people those were across their devices. Identity resolution is not part of this system, and I would rather say so than hand you a confident guess.
- Compared against your account average
- Sources, not identities
TRACE: Where the quality went. Channel level, with a line stating plainly what it cannot see.
Sample data. The rows are synthetic. The structure, the definitions, and the pipeline behind them are the real thing.
A CRO told us he logs into four platforms before every leadership meeting, picks the number he can defend, and hopes nobody asks how it was calculated. He is not missing a tool. He has four.
Every answer went through your definitions before it touched the data.
That gate is the whole mechanism. It is why two people asking the same question on a Thursday get the same number, and it is the part a screenshot cannot fake.
The gate
A question does not reach the data until it has consulted the glossary. The answer then states which definition it applied and what range it used, so a wrong assumption is visible instead of buried three queries deep.
The governed tables
Underneath every dashboard are the tables it reads. Somebody chose that dashboard's steps, filters and date range. It is a good view and it is still the wrong view the first time someone asks a question nobody anticipated. The tables answer anyway.
The artifact
You get a thing, not a paragraph. A table, a chart, a one page digest, an alert that writes itself into Slack. It lives somewhere, it refreshes, and you can send it to someone. Not an HTML file with nowhere to live.
The definitions those answers used
Four of them, published, versioned, and the same four the live dashboard reads. This is the whole gate. It is not a feature you turn on, it is a decision somebody writes down once.
- Lead
- Someone who submitted the quote request with name, address and contact details. First touch, credited to the channel of their first tracked visit.
- Qualified
- The disposition your CRM records against a lead. It is a field your team owns, not a score the reporting layer invents.
- Job Won
- A signed contract. Won revenue is attached to this step and to no other.
- Target
- The monthly and quarterly number set in the glossary, versioned there, so pacing compares against a figure nobody has to relitigate.
Same Claude. It just stops guessing.
Nobody has to learn a new tool. Your organization gets one MCP endpoint, your team adds it to Claude or Cursor with a single config, and the built-in reports and chat need no setup at all.
- One endpoint for the whole company, so the answer is the same in Claude, in Cursor, and in the built-in chat.
- Your metric definitions and your Context Warehouse sit in front of the data, so the model stops inferring what revenue means.
- No per-token usage bill from us. Your AI spend stays on the seats your team already pays for.
- Bring your own Anthropic key if you would rather the traffic run on your account.
Giving the whole company access is the governed path, not the loose one.
AI access is where most platforms get sloppy. Here it is the most controlled route into the system: the AI layer pulls only the data a question actually needs, and API traffic is not used to train anyone's models.
Questions people actually ask
What is an MCP, in plain English?
A secure connection that lets the AI tools your team already uses read your actual data, your metric definitions, and your existing reports. Instead of a chatbot that guesses, you get answers built from your source of truth, with your rules applied first.
How is this different from pointing Claude at our database?
Raw database access means the AI guesses at your business logic. It sees a column called revenue and decides what it means. Here, every question consults your metric definitions before it touches the data, and the answer states which definition it applied. Same Claude. It just stops guessing.
What happens when it does not know?
It says so. If a source is not connected, the honest answer is that the number cannot be produced yet, plus what it would take to produce it. An estimate that ends up in a board deck is worse than no answer, because by the following week it has become the number.
Do we have to use your chat interface?
No. Most teams keep using Claude or Cursor and connect through us. Your organization gets one MCP endpoint, your team adds it with a single config, and the built-in reports and chat need no setup at all.
Is this safe to give the whole company?
Programmatic access is the most governed path in the system, not the loosest. Per-application keys, single-tenant scope, rate limits, instant revocation, and the acting identity recorded on every request. The AI layer pulls only the data a question actually needs, and API traffic is not used to train anyone's models.
Who is this for?
Companies between $1M and $1B in revenue where more than one team needs numbers and exactly one person can currently produce them. If your analyst has become a chart vending machine, that is the shape.
They consistently impress by collecting the right data, organizing it into a compelling story, and creating easy-to-read, highly actionable reports.
Chris Mercer
Co-founder
Keep looking around
The Data OS
The full tour — what it is, what you get, and how engagements work.
Take a look →Why we built it
Most teams don't have a data problem — they have a trust problem.
Take a look →See an example
Four live dashboards you can filter and sort, with every number defined and queryable through your MCP.
Take a look →Who runs what
Fully managed by Vision Labs, or run by your team. Drag the blocks and set the split.
Take a look →Security & Trust
Encryption, tenant isolation, least privilege, audit — the posture IT will ask about.
Take a look →For your boss (and IT)
The forwardable one-pager that gets you sign-off, with a copy-paste email.
Take a look →Ready when you are.
Bring the four questions your team keeps failing to answer. If we cannot show you where each one would come from, that is a useful thing to learn in 30 minutes.