Yes. The repo is MIT licensed. You need Claude Code with a Claude subscription to run it, and Python 3.10 or later.

Does my data leave my laptop?

The Python tools process your data on your machine, but the prompts and the context the analyst reads go to the model provider, and an export leaves when you run one. The two hooks in .claude/settings.json write local log files only. The README covers this under "What runs on your machine".

How is it different from the Cowork plugin?

The plugin, ai-analyst-plugin, packages the same method for Claude Cowork and Claude Code as skills only, without the pipeline or the eval harness. The repo is the full system.

AI Analyst: open-source data analyst for Claude Code

One recorded run

This is from the free session Root Cause Analysis in Claude Code, recorded May 20, 2026. Hai Guan runs it on NovaMart, the synthetic online store we teach with.

What is inside

CLAUDE.md is the method. Claude Code reads it at the start of every session, and it holds ten rules, from "Frame the decision before touching data" to "Never expose credentials."

Skills are 64 markdown files that apply when they are relevant, so you rarely call one by name. docs/SKILLS.md sorts them into eight kinds:

Kind	Skills	For example
Method	7	frame the question, compare every number, pair a win with its guardrail metrics
Data	11	connect a source, profile tables, check data quality
Analysis	8	plan an analysis, forecast, run the full pipeline
Experiments and causal	4	A/B test design and readout, sample ratio check, causal methods
Checks and evals	13	reliability, trace, triangulation, score, eval
Charts and exports	8	chart style, Google Docs and Slides, Notion, Slack
Knowledge and memory	7	metric definitions, corrections, saved join patterns
Setup	6	/setup, switch dataset, hand off a long session

Pipeline agents are 40 step definitions, listed in agents/registry.yaml, that /run-pipeline runs one after another for a full analysis from question to slide deck.

Python helpers are 151 modules that hold the parts that should not vary between runs: experiment statistics, causal methods, forecasting, validation and provenance logging. In the README's words, "The model reasons; the arithmetic runs in code."

The checks are four commands. /reliability runs one question in five fresh trials and reports how often they agree. Its skill file says: "A wrong analysis can repeat perfectly." /trace-analysis follows a claim back to its query and source. /triangulation compares independent analyses that use different methods. /score-analysis ends in act, investigate, abstain or incomplete.

Evals hold 22 complete-analysis cases on NovaMart in evals/cases/public/. /eval locks each run's output before it is graded, so you can change the system, rerun and compare. The course's held-out answers are not in the repo.

Ten-minute quickstart

git clone https://github.com/ai-analyst-lab/ai-analyst.git
cd ai-analyst
pip install -e ".[dev]"
claude

You need Python 3.10 or later and Claude Code with a Claude subscription. Node.js 18 or later only matters for slides as PDF or HTML. Inside Claude Code, type /setup. It asks in four rounds, two or three questions at a time:

Your role, how technical you are with data, your team, and what kind of business your product is in.
What data you have: CSV files, a DuckDB file, a cloud warehouse, or nothing yet, in which case it offers a sample dataset.
What the company does, the two or three metrics your team cares about most, and the question you are working on now.
How much detail you want, how many charts, and how you usually share results.

Then ask a real question. The README's example:

/analyst Using the orders table, how did revenue trend last quarter
versus the one before, and where did we lose the most?
Save me a one-page brief with one chart.

Connecting Snowflake, BigQuery and the rest is on the Claude Code guide.

What it is good at, and what it is not

On September 25 we ran five data questions through the repo in fresh Claude Code sessions (the full results are on ChatGPT vs Claude for data analysis). It caught both statistical traps. It explained a Simpson's paradox in 14 seconds, and on a rigged A/B test its first line said the test was not trustworthy enough to read a lift from, because the split was 52.5/47.5 when it should be 50/50. When a prompt said orders fell, it checked first and answered that they rose 4.9 percent. Handed a CFO's count of active customers and asked for revenue per active customer, all three runs pointed out that the ratio divided revenue without cancelled orders by a count that included them.

It does not know your metric definitions until you write them down. Asked how many active customers we have right now, in four fresh sessions, three of them on the same model, it gave four numbers, from 5,526 to 14,371, each with a different definition. Every run said the project had no definition and listed the alternatives, but a number still came back. /metric-spec writes the definition, and the metric contract page covers what goes in one. In the CFO test, two of the three runs saved the handed-down definition as the official metric, and none asked first.

The same runs turned up a bug in the repo. The prompts said not to connect to Snowflake, and three of the first twelve attempts queried our warehouse anyway, read-only, because the saved settings switched the warehouse on and the connection code ignored the local dataset. The fix was merged into main on September 26: a named dataset is now honored, and an off switch set in your shell now overrides the saved setting.

BigQuery is implemented, with a setup guide, but its tests run against fake clients and we have not run it against a live project in a recorded session. Check the first few numbers against the BigQuery console, and check the README's "Your data" section for current status. For the checks to run on any answer before you share it, see how to check an AI data analyst's answer.

Where to go next

How to build an AI data analyst: the seven build steps with this repo, from install to answer key.
How to use Claude Code for data analysis: a recorded run and how to connect each data source.
ai-analyst-plugin: the same method for Claude Cowork, without the pipeline or evals.
Agentic Analytics: Build an AI Analyst: the course built on this repo. 5 weeks, 10 live sessions of 2 hours, weekly office hours, $2,500. The next cohort starts Nov 2.

The AI Analyst repo