Your agent asks the same questions either way. One of them costs more.
Everything here resolves against your own numbers — the money, the call volume, and the questions a single publisher's MCP structurally cannot reach.
Your footprint
One Skai connector
4,293 tool calls a month
$39.62
3 publisher MCPs, called directly
55,756 tool calls a month
$394.12
At 90 accounts and 300 questions a month, that is 9.9× the cost to ask the same things — and the gap widens, not narrows, as accounts are added.
12/25
questions a single publisher's own MCP can reach
With 3 publishers in your mix, 13 of the 25 questions we tested need data from more than one of them. A single publisher's own MCP tops out at 12 — not because it is slow or expensive, but because the data isn't there. One connector answers all 25.
Projected monthly cost against accounts in scope, for Skai's Reporting MCP and for calling publisher MCPs directly. At 90 accounts the projection marks $39.62 a month for Skai's Reporting MCP against $394.12 for calling publisher MCPs directly. Both curves are modeled beyond a measured 2-account baseline.
Both curves come from measured per-query baselines: five reporting workflows run twice each, against publishers whose own MCPs ship small, read-only tool schemas. Skai's scaling exponent is a real measurement — the same cross-account question asked against 1 account and against 10 under one login moved the cost only 27%.
Where the money actually goes
Not call count. Payload size. A publisher MCP with 100+ tools carries its whole schema into every single call, so a trivial question costs about what a hard one does. In our most recent round, the cheapest question that side answered all round still cost 143k tokens. Skai's cheapest cost 6k.
5.2× the tokens
For the same battery of questions, at the same tested scale.
Summed medians across every single-question task we ran, 10 runs a side. That side used fewer tool calls than Skai in this round — just far bigger ones. Every question, both sides, one scale.
- Cost per query
- The range is the publisher MCP's size and scope, not cherry-picking: 1.5× against a handful of read-only reporting tools, up to 8× against a 100+-tool one
- 1.5–8× cheaper
- Real, data-backed answers
- 810 runs, three independent judges. Completeness, not correctness — where a publisher's MCP answers, it is usually accurate
- ~2.5× more often
- Unattended runs
- A standard refresh-token login that never needs re-approval, across a batch of any length. Some publisher MCPs default to interactive browser authorization
- No re-approval
The one gap that never closes
Every number above is a snapshot. A publisher could halve their schema tomorrow and close it. This one they cannot close at all — the data for the other channels was never theirs to return.
A single publisher's own MCP
- Answers for its own channel
- Blended, cross-publisher ROAS
- “Which publisher is underperforming?”
- One reconciliation across every channel
Skai's Reporting MCP
- Answers for its own channel
- Blended, cross-publisher ROAS
- “Which publisher is underperforming?”
- One reconciliation across every channel
- Amazon Ads
- Google Ads
- Meta
Don't take the summary for it
Everything above is condensed from one benchmark. The other tab is that benchmark in full: every question we asked, what each one cost on both sides, which of them no single publisher could reach, and how the answers were graded — with the model's measured and modeled halves named separately.
Read the full benchmarkOne connector, whichever agent you already run
Claude, ChatGPT, Cursor, Gemini — the Reporting MCP behind every number on this page connects the same way to all of them.