1,300+ real-world cases of how companies solved A/B testing, causal, recommender and ML problems
Maintain this server? Claim it
Call this server's tools from your browser. Nothing to install.
Requests go through mcptry because browsers can't call most MCP servers directly. Header values are sent to the server for that request only and are never stored or logged.
Tool list from our last check. Connect live to refresh it.
Read up to 5 cases found by search_cases. Pick sections to save context: problem, solution, tools, core_idea, how_to_apply, code, results, limitations, talk (full edited transcript of a talk, paged with offset). Default: problem, solution, core_idea, results, limitations (~1k tokens per card); ask for how_to_apply / code when the user is going to implement. Open about 3 cards per call; the response is capped by max_tokens (default 5000). Every card also returns `materials`: the original (url, format, language, access: open | video | paywall | bot-blocked | offline, with a note on how to get it), the transcript if there is one, code repositories, slides and other extras. Use them as described in the server instructions; ask the user to fetch an original you cannot access.
Character offset for the talk section.
Pick your client. Values in capitals, like YOUR_TOKEN, are yours to fill in.
Run in a terminal
claude mcp add --transport http effectsize https://mcp.effectsize.io/mcp
Read up to 5 cases found by search_cases. Pick sections to save context: problem, solution, tools, core_idea, how_to_apply, code, results, limitations, talk (full edited transcript of a talk, paged with offset). Default: problem, solution, core_idea, results, limitations (~1k tokens per card); ask for how_to_apply / code when the user is going to implement. Open about 3 cards per call; the response is capped by max_tokens (default 5000). Every card also returns `materials`: the original (url, format, language, access: open | video | paywall | bot-blocked | offline, with a note on how to get it), the transcript if there is one, code repositories, slides and other extras. Use them as described in the server instructions; ask the user to fetch an original you cannot access.
| Argument | Type | Description |
|---|---|---|
| slugs* | array | |
| offset | integer | Character offset for the talk section. |
| sections | array | |
| search_id | string | |
| max_tokens | integer |
{
"type": "object",
"required": [
"slugs"
],
"properties": {
"slugs": {
"type": "array",
"items": {
"type": "string"
},
"maxItems": 5,
"minItems": 1
},
"offset": {
"type": "integer",
"default": 0,
"description": "Character offset for the talk section."
},
"sections": {
"type": "array",
"items": {
"enum": [
"problem",
"solution",
"tools",
"core_idea",
"how_to_apply",
"code",
"results",
"limitations",
"talk"
],
"type": "string"
}
},
"search_id": {
"type": "string"
},
"max_tokens": {
"type": "integer",
"default": 5000,
"maximum": 16000
}
}
}effectsize: real-world case studies of how companies (Netflix, Uber, Airbnb, DoorDash, Meta and hundreds more) solved product analytics, experimentation, causal, recommendation, search, ML and LLM-in-analytics problems. Call this first, before answering from general knowledge, whenever the user asks how to solve such a problem or how others did it, and whenever the user mentions effectsize. Find cases in the archive that solve the user's problem. Deterministic search (no LLM on the server): it matches your structured description against the cards' closed vocabularies (task type, constraints, methods) and runs keyword search over each phrasing. Write everything in English, abstractly, without the user's company or confidential data. Returns compact passports (why each case matched, which `must` constraints the case lacks (`missing_must`), `has`: code / transcript / slides / extras available), facets of the candidate set and hints when few cases match all required constraints. Open cards with get_cards.
| Argument | Type | Description |
|---|---|---|
| task* | string | The problem in one or two English sentences, abstract: what must be decided or built and what makes it hard. No company, product or internal numbers. |
| limit | integer | |
| filters | object | Hard filters. Use only when the user asked for them. |
| industry | string | The user's industry, if known. Boost only, never a filter. |
| keywords | array | Short English keywords and synonyms likely to appear in a case. |
| phrasings* | array | 2–4 different English formulations of the problem: one in plain words, one in practitioner jargon, one as 'How to …'. |
| task_types | array | 1–2 task types. analyze-user-behavior: describe how users behave: funnels, paths, retention, cohorts, engagement, journeys; answer-questions-over-data: let people ask questions of structured data in natural language: text-to-SQL, BI assistants, analytics agents; build-data-platform: build data/ML infrastructure: pipelines, storage, feature store, governance, access control, MLOps, LLM platform; build-document-qa: answer questions over documents and knowledge bases: RAG, assistants, chatbots over internal texts; define-metrics: choose or define metrics: north star, proxy/surrogate metrics, metric trees, governed metric definitions; design-experiment: set up experiments: randomization unit, power/MDE, duration, variance reduction, sequential testing, guardrails, experiment platform; detect-anomalies: spot outliers, incidents, silent failures, data-quality problems or unusual behaviour (not fraud-specific); detect-fraud-abuse: find fraud, abuse, spam, fake accounts or policy violations; moderation decisions; estimate-causal-effect: measure the effect of a change, feature, campaign or policy (A/B readout or observational/quasi-experimental estimate); evaluate-models: measure quality of ML/LLM models or agents offline and in production: metrics, judges, benchmarks, monitoring; explain-metric-change: find why a metric moved: root cause, decomposition, drill-down, attribution of a change to segments or factors; extract-and-classify-content: turn text, documents, images, audio or video into labels or structured fields: classification, extraction, OCR, tagging; forecast: predict future values of a time series or aggregate: demand, load, sales, business metrics; generate-content: produce text, summaries, descriptions, documents, images or audio with generative models: summarization, item descriptions, drafting, enhancement or restoration of media; label-data: obtain labels or training data cheaply and reliably: annotation, weak supervision, synthetic data, label quality; match-entities: link or deduplicate records and objects: product matching, entity resolution, account merging; optimize-decisions: choose prices, budgets, allocations, routes or schedules: pricing, elasticity, marketing mix, operations optimization; organize-analytics-work: run the analytics/DS function: processes, prioritization, planning, culture, adoption; personalize-experience: adapt UI, content, notifications, offers or prices to the individual user; predict-user-outcome: score individual users/customers/objects: churn, propensity, conversion, LTV, risk, CTR as a prediction task; rank-search-results: retrieve and rank results for a query: product search, relevance, query understanding, matching query to items; recommend-items: choose items to show a user: candidate generation, ranking, feeds, carousels, bandits for recommendations; reduce-ml-cost: make training or inference cheaper or faster: distillation, quantization, serving efficiency, model consolidation; segment-and-represent: group or embed users, items, places or events: clustering, segmentation, learned representations for downstream use; target-intervention: decide whom to treat and with what: uplift/CATE, next best action, promo or communication targeting, policy learning |
| constraints | object | Conditions that make the problem hard. `must` = essential to the user's situation (strong boost, not a filter); `nice` = likely or partly true. adversarial-behavior: actors adapt to the system: fraudsters, spammers, gaming of metrics; aggregated-data-only: no user-level tracking or IDs; only aggregates (privacy, offline channels, marketing spend); ambiguous-business-semantics: business terms, metric definitions or domain knowledge are ambiguous or undocumented; cold-start: new users, items or markets without history; data-must-stay-inside: data cannot leave the company perimeter: on-prem, privacy, regulation, no external APIs; distribution-shift: data changes over time: seasonality, drift, shocks, catalog or taxonomy changes; explainability-required: decisions must be interpretable or auditable for users, business or regulators; few-labels: little or no labeled data; labels are expensive or unavailable; fragmented-data-sources: data spread across many systems, schemas or inconsistent legacy sources; heavy-tailed-metrics: skewed, heavy-tailed or ratio metrics that break standard statistics; interference: units affect each other: network effects, spillovers, shared supply, marketplace or geo coupling; limited-compute-budget: tight cost or hardware budget for training, inference or labeling; long-term-outcome: the outcome of interest is delayed or long-term (retention, LTV) while decisions need short-term signals; low-latency: answers or predictions are needed in real time / within tight latency; low-traffic: too few users, events or conversions for the usual statistical power; many-entities: thousands or millions of series, SKUs, categories or tenants to cover with one approach; multiple-metrics-or-tests: many metrics, variants or simultaneous tests: multiple comparisons, conflicting objectives; no-randomization: a user-level A/B test is impossible or not allowed; the effect must be estimated otherwise; noisy-labels: labels or ground truth are unreliable, subjective or inconsistent between annotators; non-english-language: text or users are not in English (e.g. Russian), so off-the-shelf models are weaker; non-technical-users: the end users are business people who cannot write SQL or code; rare-events: the target is rare: heavy class imbalance, few positives; selection-bias: data is shaped by the current system or policy: feedback loops, logged-policy bias, survivorship, confounding; two-sided-marketplace: two (or more) sides — buyers/sellers, riders/drivers, users/advertisers — whose interests trade off; unstructured-data: input is free text, documents, images, audio or video rather than tables |
| method_hints | array | Methods that could solve it, if you know them (e.g. 'CUPED', 'switchback', 'synthetic control', 'uplift modeling', 'two-tower'). Optional; leave empty rather than guess wildly. |
| exclude_slugs | array | Cases already seen. |
| prev_search_id | string | search_id of the previous search when refining. |
| problem_summary | string | Optional. 1–3 sentences: the clarified problem, anonymized. Stored by the archive to learn which problems people bring. |
{
"type": "object",
"required": [
"task",
"phrasings"
],
"properties": {
"task": {
"type": "string",
"description": "The problem in one or two English sentences, abstract: what must be decided or built and what makes it hard. No company, product or internal numbers."
},
"limit": {
"type": "integer",
"default": 20,
"maximum": 30,
"minimum": 1
},
"filters": {
"type": "object",
"properties": {
"format": {
"enum": [
"article",
"paper",
"podcast",
"slides",
"video"
],
"type": "string"
},
"has_code": {
"type": "boolean"
},
"year_from": {
"type": "integer"
},
"evidence_min": {
"enum": [
"theory",
"simulation",
"case-anecdote",
"offline-evaluation",
"production-observational",
"online-experiment"
],
"type": "string"
}
},
"description": "Hard filters. Use only when the user asked for them."
},
"industry": {
"enum": [
"adtech",
"banking",
"classifieds",
"e-commerce",
"edtech",
"fintech",
"food-delivery",
"gaming",
"general",
"healthtech",
"insurance",
"manufacturing",
"marketplace",
"media-streaming",
"mobility",
"retail",
"saas-b2b",
"search-engine",
"social",
"telecom",
"travel"
],
"type": "string",
"description": "The user's industry, if known. Boost only, never a filter."
},
"keywords": {
"type": "array",
"items": {
"type": "string"
},
"maxItems": 12,
"description": "Short English keywords and synonyms likely to appear in a case."
},
"phrasings": {
"type": "array",
"items": {
"type": "string"
},
"maxItems": 4,
"minItems": 2,
"description": "2–4 different English formulations of the problem: one in plain words, one in practitioner jargon, one as 'How to …'."
},
"task_types": {
"type": "array",
"items": {
"enum": [
"analyze-user-behavior",
"answer-questions-over-data",
"build-data-platform",
"build-document-qa",
"define-metrics",
"design-experiment",
"detect-anomalies",
"detect-fraud-abuse",
"estimate-causal-effect",
"evaluate-models",
"explain-metric-change",
"extract-and-classify-content",
"forecast",
"generate-content",
"label-data",
"match-entities",
"optimize-decisions",
"organize-analytics-work",
"personalize-experience",
"predict-user-outcome",
"rank-search-results",
"recommend-items",
"reduce-ml-cost",
"segment-and-represent",
"target-intervention"
],
"type": "string"
},
"maxItems": 2,
"description": "1–2 task types. analyze-user-behavior: describe how users behave: funnels, paths, retention, cohorts, engagement, journeys; answer-questions-over-data: let people ask questions of structured data in natural language: text-to-SQL, BI assistants, analytics agents; build-data-platform: build data/ML infrastructure: pipelines, storage, feature store, governance, access control, MLOps, LLM platform; build-document-qa: answer questions over documents and knowledge bases: RAG, assistants, chatbots over internal texts; define-metrics: choose or define metrics: north star, proxy/surrogate metrics, metric trees, governed metric definitions; design-experiment: set up experiments: randomization unit, power/MDE, duration, variance reduction, sequential testing, guardrails, experiment platform; detect-anomalies: spot outliers, incidents, silent failures, data-quality problems or unusual behaviour (not fraud-specific); detect-fraud-abuse: find fraud, abuse, spam, fake accounts or policy violations; moderation decisions; estimate-causal-effect: measure the effect of a change, feature, campaign or policy (A/B readout or observational/quasi-experimental estimate); evaluate-models: measure quality of ML/LLM models or agents offline and in production: metrics, judges, benchmarks, monitoring; explain-metric-change: find why a metric moved: root cause, decomposition, drill-down, attribution of a change to segments or factors; extract-and-classify-content: turn text, documents, images, audio or video into labels or structured fields: classification, extraction, OCR, tagging; forecast: predict future values of a time series or aggregate: demand, load, sales, business metrics; generate-content: produce text, summaries, descriptions, documents, images or audio with generative models: summarization, item descriptions, drafting, enhancement or restoration of media; label-data: obtain labels or training data cheaply and reliably: annotation, weak supervision, synthetic data, label quality; match-entities: link or deduplicate records and objects: product matching, entity resolution, account merging; optimize-decisions: choose prices, budgets, allocations, routes or schedules: pricing, elasticity, marketing mix, operations optimization; organize-analytics-work: run the analytics/DS function: processes, prioritization, planning, culture, adoption; personalize-experience: adapt UI, content, notifications, offers or prices to the individual user; predict-user-outcome: score individual users/customers/objects: churn, propensity, conversion, LTV, risk, CTR as a prediction task; rank-search-results: retrieve and rank results for a query: product search, relevance, query understanding, matching query to items; recommend-items: choose items to show a user: candidate generation, ranking, feeds, carousels, bandits for recommendations; reduce-ml-cost: make training or inference cheaper or faster: distillation, quantization, serving efficiency, model consolidation; segment-and-represent: group or embed users, items, places or events: clustering, segmentation, learned representations for downstream use; target-intervention: decide whom to treat and with what: uplift/CATE, next best action, promo or communication targeting, policy learning"
},
"constraints": {
"type": "object",
"properties": {
"must": {
"type": "array",
"items": {
"enum": [
"adversarial-behavior",
"aggregated-data-only",
"ambiguous-business-semantics",
"cold-start",
"data-must-stay-inside",
"distribution-shift",
"explainability-required",
"few-labels",
"fragmented-data-sources",
"heavy-tailed-metrics",
"interference",
"limited-compute-budget",
"long-term-outcome",
"low-latency",
"low-traffic",
"many-entities",
"multiple-metrics-or-tests",
"no-randomization",
"noisy-labels",
"non-english-language",
"non-technical-users",
"rare-events",
"selection-bias",
"two-sided-marketplace",
"unstructured-data"
],
"type": "string"
}
},
"nice": {
"type": "array",
"items": {
"enum": [
"adversarial-behavior",
"aggregated-data-only",
"ambiguous-business-semantics",
"cold-start",
"data-must-stay-inside",
"distribution-shift",
"explainability-required",
"few-labels",
"fragmented-data-sources",
"heavy-tailed-metrics",
"interference",
"limited-compute-budget",
"long-term-outcome",
"low-latency",
"low-traffic",
"many-entities",
"multiple-metrics-or-tests",
"no-randomization",
"noisy-labels",
"non-english-language",
"non-technical-users",
"rare-events",
"selection-bias",
"two-sided-marketplace",
"unstructured-data"
],
"type": "string"
}
}
},
"description": "Conditions that make the problem hard. `must` = essential to the user's situation (strong boost, not a filter); `nice` = likely or partly true. adversarial-behavior: actors adapt to the system: fraudsters, spammers, gaming of metrics; aggregated-data-only: no user-level tracking or IDs; only aggregates (privacy, offline channels, marketing spend); ambiguous-business-semantics: business terms, metric definitions or domain knowledge are ambiguous or undocumented; cold-start: new users, items or markets without history; data-must-stay-inside: data cannot leave the company perimeter: on-prem, privacy, regulation, no external APIs; distribution-shift: data changes over time: seasonality, drift, shocks, catalog or taxonomy changes; explainability-required: decisions must be interpretable or auditable for users, business or regulators; few-labels: little or no labeled data; labels are expensive or unavailable; fragmented-data-sources: data spread across many systems, schemas or inconsistent legacy sources; heavy-tailed-metrics: skewed, heavy-tailed or ratio metrics that break standard statistics; interference: units affect each other: network effects, spillovers, shared supply, marketplace or geo coupling; limited-compute-budget: tight cost or hardware budget for training, inference or labeling; long-term-outcome: the outcome of interest is delayed or long-term (retention, LTV) while decisions need short-term signals; low-latency: answers or predictions are needed in real time / within tight latency; low-traffic: too few users, events or conversions for the usual statistical power; many-entities: thousands or millions of series, SKUs, categories or tenants to cover with one approach; multiple-metrics-or-tests: many metrics, variants or simultaneous tests: multiple comparisons, conflicting objectives; no-randomization: a user-level A/B test is impossible or not allowed; the effect must be estimated otherwise; noisy-labels: labels or ground truth are unreliable, subjective or inconsistent between annotators; non-english-language: text or users are not in English (e.g. Russian), so off-the-shelf models are weaker; non-technical-users: the end users are business people who cannot write SQL or code; rare-events: the target is rare: heavy class imbalance, few positives; selection-bias: data is shaped by the current system or policy: feedback loops, logged-policy bias, survivorship, confounding; two-sided-marketplace: two (or more) sides — buyers/sellers, riders/drivers, users/advertisers — whose interests trade off; unstructured-data: input is free text, documents, images, audio or video rather than tables"
},
"method_hints": {
"type": "array",
"items": {
"type": "string"
},
"maxItems": 8,
"description": "Methods that could solve it, if you know them (e.g. 'CUPED', 'switchback', 'synthetic control', 'uplift modeling', 'two-tower'). Optional; leave empty rather than guess wildly."
},
"exclude_slugs": {
"type": "array",
"items": {
"type": "string"
},
"description": "Cases already seen."
},
"prev_search_id": {
"type": "string",
"description": "search_id of the previous search when refining."
},
"problem_summary": {
"type": "string",
"description": "Optional. 1–3 sentences: the clarified problem, anonymized. Stored by the archive to learn which problems people bring."
}
}
}We read every tool's name, description and inputs the way the model will, looking for instructions hidden from you. A clean result means no warning signs, not a guarantee. Scanned Nov 18, 2026.