Ask your assistant a question about a spreadsheet today and it will write a little program on the spot. Usually that is fine. When it is not, you cannot tell — it might have used the wrong test, or graded its prediction on the rows it learned from. The answer looks equally confident either way.
meelu-analytics-mcp sits between your assistant and your data and takes those judgement calls out of its hands. Forty-five tools, the method chosen from your actual data, and a trust rating on every answer — including a refusal when the data cannot support the question.
One command from the README installs the server, asks which assistant you use and writes the config for you. It works with 10 assistants — Claude Code, Claude Desktop, OpenAI Codex, Cursor and more. Free and MIT licensed: no account, no key.
Use cases
These are not separate products or plugins. It is one server — install it once and every one of these is available to your assistant.
Work out what predicts a customer leaving — and who is at risk now
View use caseProject revenue forward with trend, seasonality and honest error bars
View use caseFind out what a customer is worth, and which traits predict it
View use caseTurn a Stripe export into churn drivers, forecasts and cohorts
View use caseSee which signup months held up, and find the month things changed
View use caseRFM segments and natural groupings found in your data
View use caseCapabilities
Built on DuckDB, scikit-learn, statsmodels, SHAP and DoWhy. All 45 tools, by category.
Why it is built this way
Comparing groups against a number usually calls for one particular test — unless your data breaks that test's assumptions, in which case it needs a different one, and the difference changes the answer. That check runs every time, and what it found is reported with the result.
Too few rows to be meaningful. A prediction that would be graded on the rows it learned from. A cause-and-effect claim that fails its own sanity check. In each case you get a plain explanation instead of a number, because a refusal tells you something true and a confident wrong answer does not.
Segments, predictions and any new columns you create are saved back alongside your data, so each question builds on the last instead of starting over. It survives shutting down and coming back tomorrow.
There is no paid tier, no row cap and no account. The server runs on your machine and reads your files from your disk. If that changes, it will change on this page first — for now, the honest description of the business model is that there is not one.
The one caveat worth stating plainly: cloud assistants such as Claude on the web cannot start a program on your laptop, so connecting them means exposing the server over the network and your data does leave your machine. For local files, use a desktop assistant.
FAQ
CSV files, today. One row per thing you care about — customers, orders, subscriptions, events. Name several related files in the same message and the engine works out how they join. Excel, Parquet and JSON are on the roadmap; for now, save as CSV first.
It runs on your own machine and reads files from your own disk, so nothing is uploaded anywhere. One honest exception, which the README states plainly: if you expose the server over HTTP so a cloud assistant like Claude on the web can reach it, your data does leave your machine. For local files, use one of the desktop assistants.
The installer detects and configures 10: Claude Code, Claude Desktop, OpenAI Codex, Cursor, Windsurf, VS Code (Copilot), Gemini CLI, Grok CLI, OpenCode, Zed. Anything else that reads a standard MCP config works too — the installer can print the config block for you to paste in.
No. You ask in plain English and your assistant picks the tools. Knowing some statistics helps you interrogate the answer, which is worth doing — but the three follow-ups that matter (which method, how much to trust it, is it causal) are questions anyone can ask.
Yes. MIT licensed, no account, no API key, no tier. It is built on DuckDB, scikit-learn, statsmodels, SHAP and DoWhy, which are free as well. There is no paid version of this to upsell you to.
Comfortably into the hundreds of thousands of rows on a normal laptop. Very large files may currently be refused rather than sampled — graceful handling of those is explicitly on the roadmap.
Dashboards, scheduled reports, and connecting to your other systems. It is the analysis engine; what you do with the answer is up to you. That is a stated design decision rather than a gap waiting to be filled.
One command from the README installs the server, asks which assistant you use and writes the config for you. It works with 10 assistants — Claude Code, Claude Desktop, OpenAI Codex, Cursor and more. Free and MIT licensed: no account, no key.