
effectsize
io.github.vzhyvov/effectsize
1,300+ real-world cases of how companies solved A/B testing, causal, recommender and ML problems
Documentation
effectsize skill and MCP server
Give your AI agent 1,300+ real-world case studies of how companies solved product analytics, A/B testing, causal inference, recommendation, search, fraud, forecasting, ML and LLM-in-analytics problems. Ask a question, and the agent answers with what actually worked at Netflix, Airbnb, Uber, DoorDash, Meta and hundreds of other companies, with a link to each case and its original post or talk.
Cases come from effectsize.io: company engineering blogs, papers and conference talks (many not in English), each reviewed and turned into a short English card.
Try it without signing in
Add this MCP server to any client that supports remote MCP (Streamable HTTP):
https://mcp.effectsize.io/mcp/public
Search works fully; each case returns its problem, solution and tools. For full cards (how to apply, results, limitations, code, talk transcripts) use the main endpoint below with a free sign-in.
Install
Claude Code — the plugin adds the skill and the MCP server:
/plugin marketplace add vzhyvov/effectsize-skill
/plugin install effectsize@effectsize
Or just the MCP server:
claude mcp add --transport http --scope user effectsize https://mcp.effectsize.io/mcp
Then /mcp → effectsize → Authenticate (Google or an email code).
Cursor — mcp.json:
{ "mcpServers": { "effectsize": { "url": "https://mcp.effectsize.io/mcp" } } }
Claude, ChatGPT and other clients — step-by-step guide: https://effectsize.io/mcp/
Other agents that read skills — copy skills/effectsize/SKILL.md into your
client's skills folder.
Ask
Start with "Use effectsize:" so the agent searches the cases instead of answering from memory:
- Use effectsize: we can't randomize users for a pricing change. How did others measure the effect?
- Use effectsize: our A/B tests take too long on low traffic. What variance reduction worked in practice?
- Use effectsize: how do marketplaces evaluate and improve search ranking?
Tools
| Tool | What it does |
|---|---|
search_cases | Finds cases for an abstract description of the problem, its constraints and candidate methods. Deterministic search, no LLM on the server. |
get_cards | Reads up to 5 cases: problem, solution, tools (public); core idea, how to apply, results, limitations, code, talk transcript (signed in). |
Server card: https://mcp.effectsize.io/.well-known/mcp/server-card.json
Privacy
Search arguments are logged to improve coverage, with a pseudonymous user id; don't send company names or confidential data (the agent is told to keep queries abstract). Details: https://effectsize.io/privacy/
License
MIT for the files in this repository. Case cards on effectsize.io summarize other people's work; originals are linked.
No installation packages available.