The agent checks every figure at the source before it reports.
AI agents miscalculate, misread fields and fill gaps with estimates. The Data Hub attaches the meaning of the fields to every result. Before answering, the agent checks its figures against a fresh query to the vendor. We don't know of another MCP hub that re-checks its agents' figures.
- Definitions for every field in a result
- Each figure checked against the source
- Required by default, read calls only
- Sessions, GA4, Oct 1–2718,420confirmed
- Google Ads cost, Octobercorrect: €3,912.40 (cost_micros ÷ 1,000,000)€4,312deviates
- New deals from paid social14confirmed
- Emails read, InstantlyInstantly counts opens, not emails read61%not checkable
One wrong number costs more than one report
The agent reads ad spend in micros and reports it as dollars. It treats email opens as reads, adds up users across several days or sums deal amounts in different currencies. The number looks plausible and ends up in your forecast call or board deck. Once someone spots the error, every number in that report is in doubt, including the right ones.
What the Data Hub does.
Field definitions on every result, and every figure checked against the source before it goes into a report.
- 01
Field definitions for every source
For GA4, Search Console, Google Ads, LinkedIn Ads, Meta Ads, HubSpot, Salesforce, Instantly, HeyReach, Clay, Lusha, SalesViewer, Fibbler, Notion, ClickUp and the cross-source reports, the Data Hub describes the standard fields: what they mean, which unit, currency and time zone they use, and where the usual traps are. When a field means something different depending on the endpoint, the definition is per tool.
- Google Ads: fields ending in “micros” are the amount in account currency times 1,000,000
- GA4: user counts are deduplicated per row and can't be added up across days
- Instantly: opens are unreliable because mail privacy features load the tracking pixel
- HubSpot: dealstage is an internal ID, open deals can have a close date in the past, and you sum amounts in several currencies via amount_in_home_currency
- Salesforce: Amount isn't weighted; ExpectedRevenue is Amount times Probability
- HubSpot: a contact's first and latest traffic source live in different fields
- 02
Data context on every result
The hub appends a short block to every read result: source, account, fetch time and the definitions of exactly the fields that appear in it, plus notes such as “don't add up user counts across days”. That puts the explanation right next to the data. The revprime__field_definitions tool returns all definitions as well.
- 03
The fact check: revprime__verify_claims
Before answering, the agent submits its figures, each with the call it took them from. The hub runs the query against the vendor again and returns one of three verdicts per figure. A single check covers up to 30 figures from up to 10 calls.
- “Confirmed”: the figure matches the source. Rounded values count as long as they match.
- “Mismatch”: the hub returns the actual value, and the agent is told to correct it before answering.
- “Unverifiable”: the agent is told to say so or leave the figure out.
- Percentages and micros are converted with a factor the agent passes along.
- 04
Required by default
The agent's instructions require the check before any statement that contains figures. Every read result reminds it, and every built-in playbook ends with it. An admin can switch the check to “recommended only” per MCP server, for example on a server that only maintains records.
- 05
Safe, and no copy of your data
The check only repeats read calls of tools enabled on that server. It never repeats a change. It counts toward the limits, shows up in the activity log and stores nothing: the agent only gets the result of the comparison.
What you should know.
- Whether the agent actually runs the check is up to the AI client. The hub asks for it at every point it controls, and the agent is told to state at the end of a report what was checked.
- Only figures from read calls can be checked. Vendors may still update the last few days' values between the first query and the check; the hub points that out when it happens.
- Tag Manager and custom REST connectors have no field definitions.
Frequently asked questions.
›What happens when a figure is wrong?
The agent gets “mismatch” along with the actual value. Its instructions say to correct every mismatch before answering and to flag or drop figures that couldn't be verified.
›Can the check change anything in my tools?
No. It only repeats read calls, and only for tools already enabled on the MCP server. Those calls count toward the limits and appear in the log.
›Can I turn the check off?
An admin can set it to “recommended only” per MCP server. Results then still suggest the check but no longer require it. Platform admins set the default in the settings.
›How do I know the figures were checked?
The agent is told to say at the end of a report which figures it checked. Every check is also recorded as a call in your business unit's activity log.
- ReportsAd spend from Google Ads, LinkedIn Ads and Meta Ads against deals and revenue in HubSpot, in a single call.
- Templates & playbooksReady-made MCP servers for reporting, CRM, outbound and ads, plus playbooks that show up as commands in your AI client.
- Guardrails & approvalsRules, approvals and limits for agents that write to your CRM and ad accounts.
See it with your own tools.
In 30 minutes, we'll show you the Data Hub with your own tools and your use case.