SpendCanary MCP¶
Ask an assistant what a Google Ads budget really costs and it will give you a number it invented. The daily budget is not the daily ceiling, the monthly ceiling is not thirty times the daily budget, and a search terms report is not a list of keywords you bid on. SpendCanary's MCP server answers those three questions from formulas rather than from prose, and it never touches your ad account, so the answer is the same whether you are signed in or not.
This is the cleanest shape a tool server can take: keyless, stateless, no account, no stored input, with the arithmetic printed next to the answer. Four of its eight tools were driven this pass, and two deliberately wrong calls returned schema errors that name the exact argument the tool wanted.
Server type: remote Streamable HTTP, stateless (no Mcp-Session-Id issued)
Endpoint: https://spendcanary.com/mcp
Server: spendcanary-free-tools, version 1.1.0
Protocol: 2025-06-18
Auth: none (no key, no account, no OAuth)
Tools: 8
Repository: github.com/spendcanary/mcp-free-tools (0 stars, created 2026-11-15)
Listed on: mcpservers.org, sitemap shard servers/12.xml
The evidence tier¶
Keyless and driven. initialize answered 200 as spendcanary-free-tools 1.1.0 on protocol 2025-06-18, with no session identifier returned and no credential at any point, which is the stateless form. tools/list returned eight tools with full input schemas, and four of them were called:
cpcwith spend 1,200 and clicks 400 answered $3.00 per click and echoed both inputs back.roaswith revenue 10,000 and ad spend 2,500 answered 4.00x (400%) and prompted formargin_percentto get a break-even figure.daily_budget_max_spendwith a 100 daily budget answered 200 in one day and 3,040 in a month, naming the 2x daily and 30.4x monthly rules.wasted_spend_from_search_termson a three-row CSV answered 47% of spend (127.60 of 269.60) across 2 zero-conversion terms and named both, with clicks beside each.
The schema is strict, and that is useful. roas called with sales returned revenue: Invalid input: expected number, received undefined, and search_partners_cost called with search_cost and partners_cost returned the four names it actually wants: google_search_cost, google_search_conversions, search_partners_cost, search_partners_conversions. An agent that guesses argument names gets a named refusal rather than a wrong number.
One integration trap, measured. A POST with no Origin header was refused 403 from this host, and the same POST with Origin: https://spendcanary.com answered 200. Any client that strips Origin has to be given one, or the connector looks broken when it is not.
Tools & Capabilities¶
| Tool | Arguments it wants | What it answers |
|---|---|---|
check_ad_copy |
platform (google, meta), ad text |
Character limits and common rejection reasons, per line |
daily_budget_max_spend |
daily_budget |
The 2x one-day ceiling and the 30.4x monthly ceiling |
roas |
revenue, ad_spend, optional margin_percent |
Return on ad spend, with break-even ROAS and profit when a margin is given |
cpc |
any two of spend, clicks, cpc |
Cost per click, plus CPM and click-through rate when impressions are given |
search_partners_cost |
google_search_cost, google_search_conversions, search_partners_cost, search_partners_conversions |
Cost per conversion on Search against Search Partners, and what partners add monthly |
wasted_spend_from_search_terms |
csv (a search terms report) |
Spend on zero-conversion terms, the costliest terms, and the share of total |
check_landing_page |
a public URL | Mobile and desktop load, redirects, a way to call or buy, the Google tag |
google_ads_problem_trends |
problem, window | How often each wasted-spend problem appears in public Reddit posts |
Every answer opens with a one-line summary, then the detail, then exactly one link to the matching calculator or guide on the vendor's site, tagged with utm_source=mcp.
Installation¶
{
"mcpServers": {
"spendcanary": { "url": "https://spendcanary.com/mcp" }
}
}
Claude Code: claude mcp add --transport http spendcanary https://spendcanary.com/mcp. In a client that asks for authentication, choose no sign-in; the server has no auth path to configure.
Why This Matters for Operators¶
Paid search is the one budget line where a rounding error compounds monthly. Three of these tools answer questions that normally require either an agency or a spreadsheet: what the real monthly ceiling is, what share of spend went to searches that never converted, and what a cost-per-conversion comparison between Search and Partners implies. Because nothing is stored and no ad account is read, the answers can be pulled into a workflow that touches account data elsewhere without turning this server into a second copy of it.
Questions operators ask¶
Does it read my Google Ads account?¶
No. The instruction block states it plainly: the calculators work on numbers and text you provide, and the only tool with any outside input is the Reddit trend counter, which reads the vendor's own public counts. There is no OAuth flow to grant.
Is the arithmetic checkable by hand?¶
Yes, which is why the numbers above are quotable. 1,200 divided by 400 is 3.00. 10,000 divided by 2,500 is 4.00. A 100 budget at twice the daily rate is 200, and at 30.4 times is 3,040. The wasted-spend answer reconciles both ways: 86.40 plus 41.20 is 127.60, and 127.60 over 269.60 is 47%.
What does it not do?¶
It does not manage campaigns, change bids, read spend, or check a live account. It also does not verify that a search terms CSV came from Google rather than a spreadsheet, so the answer is only as good as the export behind it.
Related¶
- Google Ads MCP - the campaign-side surface this calculator set does not replace
- Markifact Google Ads MCP - approval-gated Google Ads writes
- Ryze Google Ads MCP - hosted ads reporting
Integration with CorpusIQ¶
CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.
This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.