Local data lab for Claude

Claude thinks.
Prism executes.

29 tested tools. Your data stays on disk — not in chat. Every step logged, versioned, and exportable.

Download for Mac Manual install
Data never in chatVersion historyML guardrailsRuns locally
Claude · Prism MCP
Prep sales.csv for ML — check quality, fix categories, split 80/20, export pipeline
→ load_dataset("sales.csv")
→ detect_data_quality_issues()
→ harmonize_categorical_values(...)
→ train_test_split(test_size=0.2)
→ export_pipeline_config()
→ list_versions()
Pipeline exported · 5 versions · test set locked

The difference

Why not just Claude?

Claude can write pandas. Prism gives you a stateful lab with undo, guardrails, and exportable pipelines.

Data in context
Rows pasted or re-loaded each turn
Stays in server memory
Approach each session
New script, may differ
Same 29 tools every time
Undo / audit trail
None
list_versions, rollback_version, diff_versions
ML safety
Only if you prompt for it
validate_action, split immutability, cardinality guards
Reproducibility
Copy code from chat
export_pipeline_config
Privacy
Depends on product
100% local on Mac

Starter prompts

Ask Claude things like

Natural language in, structured data work out. No boilerplate scripts.

Find outliers in my sales CSV

Normalize messy category names before encoding

Split for ML and export the pipeline config

What quality issues does this dataset have?

On your Mac

Feels native in Claude

Ask in plain English. Prism runs the pipeline on your Mac — Claude never sees your rows.

claude.ai

Your data, ready for Claude

Prism has 29 tools — just ask Claude
Clean my sales CSV and find outliers Sonnet · Prism MCP
Load dataset Quality check Remove outliers Train/test split Export CSV

Getting started

How it works

From install to first analysis in under five minutes. Prism shows what it automates and what you still do manually.

01

Before you start

DMG → Applications, Python 3.10+, and Claude Desktop installed.

02

One click in Prism

Open Prism, review the checklist, click Set up Prism — progress and notifications guide you.

03

Finish in Claude

Quit Claude (Cmd+Q), reopen it, and paste the starter prompt.

Why Prism

A data lab, not a chat script

Purpose-built tools with guardrails Claude won't enforce by default.

Your data never enters chat

load_dataset reads from disk. Only summaries flow back to Claude — not your rows.

load_dataset · peek_dataset_metadata

Every step is reversible

Version snapshots after each transform. Roll back or diff any two states.

list_versions · rollback_version · diff_versions

ML mistakes caught before they run

Dry-run validation, train/test split locks the test set, cardinality guards on encoding.

validate_action · train_test_split

Export a real pipeline

Download every operation as JSON or YAML. Reproduce the exact cleaning flow later.

export_pipeline_config · generate_preprocessing_report

29
MCP tools
18
Test suites
8
Encoding methods

Claude thinks. Prism executes.

Install on Mac, connect Claude, and run your first pipeline.

Download for Mac