Spillmark

npx add-mcp 'https://spillmark.com/mcp'

Spillmark for AI agents

Connect an AI assistant or a script to Spillmark over MCP or a plain REST API. Anyone can use the free calculators. A brand's own agent can read its measured lift, next test plan, locked predictions and recommendations with a read-only key.

Server URL

https://spillmark.com/mcp

Streamable HTTP. No key needed for the public tools; add a brand key as a bearer token for the brand tools.

Connect

Claude and ChatGPT: add a custom connector with the server URL above (public tools).

Cursor: add to .cursor/mcp.json. Leave out headers for the public tools.

{
  "mcpServers": {
    "spillmark": {
      "url": "https://spillmark.com/mcp",
      "headers": { "Authorization": "Bearer sm_…" }
    }
  }
}

Claude Code:

claude mcp add --transport http spillmark https://spillmark.com/mcp --header "Authorization: Bearer sm_…"

REST: the same tools at https://spillmark.com/api/v1, which lists each tool with its input schema.

curl -X POST https://spillmark.com/api/v1/run_calculator \
  -H "Content-Type: application/json" \
  -d '{"slug": "holdout-test-calculator", "inputs": {"regions": 30, "lift": 15}}'

curl https://spillmark.com/api/v1/get_lift -H "Authorization: Bearer sm_…"

Public tools

No key. Up to 300 calls an hour from one address. The calculators are the same ones at /tools, and each result links back to the calculator with the same inputs.

• list_calculators Spillmark's free Amazon and Meta ads calculators (holdout test size, brand-size MDE, geo holdout states, ad waste, Meta ROAS reality check, incremental ROAS, break-even ROAS, ACoS, TACoS, Amazon fees and profit, LTV and CAC), with each one's inputs and example values. Use run_calculator to run one.
• run_calculator Runs one calculator from list_calculators on your numbers and returns the headline result, the supporting figures, a one-sentence reading, and a link that opens the calculator with the same inputs. Missing inputs use the calculator's example values. Enter percents as points (15 means 15%), choices as the option index.
• how_lift_is_calculated Spillmark's method in brief: how a geo holdout test measures the sales Meta ads cause on Amazon, with links to the full method and a sample report on made-up numbers.

Size a test before you connect

A Claude skill that runs the holdout calculator on your own order counts. It does not read a brand, and it does not need a key. In Claude, add the file as a skill. With the Spillmark server connected, ask it to size a test.

Download the holdout skill

Brand tools

Need a key from Settings → Agent access in the dashboard. Brand owners create keys; each one reads one brand. The key decides the brand, so no tool takes a brand name, and one brand's key can't reach another's results.

• get_brand The brand this API key reads for: its name, today's date in the brand's time zone, which data sources are connected and when each last synced, and a dashboard link.
• get_lift What this brand's Meta ads cause on Amazon (and Shopify, when tested): each holdout test's lift and extra orders a day with 95% ranges and its locked design hash, the ongoing Meta→Amazon effect estimate against break-even, and incremental CAC from the latest completed test. Null fields mean there's no result yet.
• get_test_plan The next holdout test Spillmark recommends for this brand: start and end dates, which states keep ads on and which pause, the options by holdout size with what each costs and the smallest lift it can detect, and any sales events or stock-outs overlapping the window.
• get_track_record Every prediction Spillmark locked for this brand (the change, the predicted figures with 80% ranges, the actual result once scored, and whether the stored file still matches its lock hash), how many landed in range, and whether the brand's hash-chained ledger verifies.
• get_ledger_export This brand's whole ledger (pre-registered designs, results, locked predictions and their scores) as a self-contained bundle with its hash chain, timestamps and the steps to verify it independently. Large; use get_track_record for a summary.
• list_recommendations Changes Spillmark has recommended for this brand, each with why, the data used, its locked prediction, guardrail checks and how to make it, plus 7- and 14-day results for changes the brand made. The brand approves and makes every change; this only reads.
• plan_the_test The next holdout test in one answer: dates, the recommended holdout, the smallest lift it can see, what it costs with a 95% range, and the link to lock it. Says too uncertain to call when the design cannot see break-even.
• read_the_result The latest measured lift in one sentence: the estimate, the 95% range, and too uncertain to call when that range spans zero or break-even. Links the test on the log.
• score_the_last_change The last scored prediction: what was predicted, the 80% range, what happened, and whether it was a miss, plus how many scored figures landed in range. Links the log.
• read_the_brief What a scheduled check would say: a locked test that is starting or running, the latest result with its range, a recommendation waiting for a yes, and stock split into overdue, order soon, and overstock. Nothing is sent.
• read_the_account Last 30 full days in one answer: revenue, contribution margin after fees, product cost and ad spend, Meta spend and Amazon ad charges kept separate, net profit, TACoS, and stock split into overdue, order soon, and overstock. Totals only. Amazon ad spend is Seller Central advertising charges, not an Ads API report.

Skills for a connected brand

Each file calls one brand tool and repeats its answer. Add it in Claude as a skill, with the server connected and a brand key. If the tool asks for a key, create one under Settings → Agent access.

• Plan the test
• Read the result
• Score the last change
• Read the brief
• Read the account

What agents get, and what they don't

• Results, not raw data: lift with its 95% range, the test's locked design, plans and predictions, and last-30-day profit after fees, product cost and ad spend. No order-level, customer or ad-account records.
• Read only. No tool changes ads, budgets, orders or settings. Spillmark recommends; the brand makes every change.
• Every number keeps its uncertainty, and predictions show misses as well as hits.
• Owners see which tool each key called and when, and can revoke a key at any time. Calls log the tool name, not what it returned.
• A key stops working when it's revoked or its creator loses access to the brand.

How the numbers are made: See exactly how lift is calculated.