effectsize MCP¶
The most expensive analytics mistake is not a wrong calculation. It is solving a problem that another company already solved twice and wrote down.
An engineer designing an experiment has read maybe five of the three hundred articles that answer their exact question. The knowledge exists - in engineering blogs, conference talks, papers - and it is unfindable, because the search terms that describe the problem are not the words the companies used.
effectsize is that archive turned into a retrieval surface. Every entry is a reviewed card with a problem, a solution, a method, the results, the limitations and a link to the original. The design decision worth noting is that the retrieval is not a text search: the model is asked to describe the shape of the problem, and the results come back with an explicit statement of whether the match is close or merely adjacent.
Server type: remote Streamable HTTP (hosted, two routes)
Endpoint: https://mcp.effectsize.io/mcp/public (no sign-in)
https://mcp.effectsize.io/mcp (OAuth, full cards)
Server: effectsize-cases-public, version v0.4
Protocol: 2025-06-18
Auth: none on /public (rate-limited per IP)
Tools: 2 (search_cases, get_cards)
Ranker: v0.4
Website: https://effectsize.io
Listed on: the mcp.so submission queue as #7582
The evidence tier¶
Keyless public route, driven. initialize answered 200 with no credential and returned an unusually long instructions block that is itself the product brief: search the archive before answering from general knowledge, describe the decision and the constraint rather than the company, read the passports, judge relevance yourself, and tell the user when the archive has no close case instead of citing weak matches.
tools/list returned two schemas. Both were driven: two search_cases calls on different task types, one get_cards call with two real slugs, and one get_cards call with an invented slug to measure the refusal. Five of the six calls returned data.
The full card is behind a free sign-in - the public route returns problem, solution and tools only. No sign-in was performed, no OAuth flow was started, and nothing below quotes a section the public route withheld.
The search takes a problem description, not keywords¶
search_cases requires a task sentence and a phrasings array, with an optional task_types filter over a fixed vocabulary. The instruction is explicit that the input must be abstract - no company, no product, no internal numbers.
A checkout test where the two arms contaminate each other:
task "measure whether a new checkout flow increased completed orders,
but test and control users share the same household and
promotions leak between them"
types design-experiment
full_matches 104 candidates 365
search_id a641c839dd6b snapshot e6945c5405c47d74
The reply is a ranked candidate list with a structure designed to be judged rather than trusted. Each result carries a tldr, a how line phrased as the question it answers, an evidence grade, an adjacent slug, and a why object naming the task type it matched. The top three for that query:
| Company | Year | Case | Evidence grade |
|---|---|---|---|
| Apple | 2023 | Sample-size calculations for A/B testing, novel extensions | simulation |
| Lyft | 2022 | Challenges in experimentation | case-anecdote |
| Spotify | 2023 | Sequential testing with longitudinal data, the peeking problem | simulation |
A different task type, a different result set, and the same snapshot identifier:
task "forecast weekly demand for a new market with only eight weeks
of history and a promotion calendar"
types forecast
full_matches 52 candidates 306
search_id 0e4da0bbc879 snapshot e6945c5405c47d74
Top of that list: Uber on forecasting airport demand and time-to-request, Yandex Lavka on reconciling forecasts across teams, Otto on dynamic pricing, Avito on forecasting the effect of a paid listing promotion. The snapshot is identical across both calls, which means the archive is versioned and two agents querying the same day are answering from the same corpus.
What does a card actually contain?¶
get_cards takes up to five slugs and a section list. The Lyft card's problem section, returned on the public route, is the reason this server is worth a guide:
Lyft tests every product change, but a standard user-level A/B test fails in a ride-sharing marketplace. Riders and drivers share one pool of supply, so one user's treatment changes what another user gets. The independence assumption between groups breaks.
That is a real methodological answer with the mechanism named, and its solution section names the remedy - choose the randomisation unit to fit the interference, then wrap it in process: pre-registered hypothesis, Benjamini-Hochberg correction across many primary metrics, an experiment scheduler, a decision log.
The Apple card answers a different failure: randomise users but compute the metric per session or per event, and the rows of one user are correlated, so the standard independent-samples formula understates the variance, asks for too few observations, and leaves the test underpowered.
Each card also carries a materials block declaring what exists and whether it can be fetched. Lyft's original is marked access: bot-blocked with a note to ask the user to paste it; Apple's paper is marked open. That distinction is the difference between a citation and a dead end, and stating it up front saves the fetch.
The refusal is a success envelope¶
An invented slug came back like this:
{"cards": [{"slug": "definitely-not-a-real-slug-xyz",
"error": "unknown slug", "materials": {}}],
"approx_tokens": 0}
HTTP 200, isError false, and the failure carried per card. A client that renders only the content block and not the inline error will show an empty result as if it were an answer. Callers should check each card object for an error key before treating the batch as complete.
Where it fits¶
This is a research surface, not a data source. It does not hold your metrics and it will not compute your experiment. What it does is put two hundred reviewed cases in front of the model at the moment a question is being designed, with an explicit confidence signal and an honest "no close case" state.
For an operator, the value is narrow and real: the question "how have other companies measured this" gets an answer with a company, a year, a method and a link, instead of a confident generalisation.
Is the archive behind a paywall?¶
No, but it is split. search_cases and the problem/solution/tools sections of each card are free on the /public route, rate-limited per IP. The full card - core idea, how to apply, results, limitations, code, talk transcripts - needs a free sign-in through OAuth on /mcp, or an account on the website. Nothing in this page relies on a gated section.
Does it train on the query?¶
The server says the search arguments are stored to improve coverage and asks the caller not to include company names, people, internal figures or other confidential details. Treat the task sentence as a public input.
Sources¶
- Public endpoint: https://mcp.effectsize.io/mcp/public
- Full endpoint: https://mcp.effectsize.io/mcp
- Homepage: https://effectsize.io
- Skill and source: https://github.com/vzhyvov/effectsize-skill
- Submission: the mcp.so queue as #7582 on 2026-11-18
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.