Mostly Right docs

Mostly Right helps you discover, build and maintain datasets for machine learning. Browse published Parquet datasets, or let an agent find sources, configure a pipeline and keep your own dataset updated.

What a dataset is here

A dataset contains one or more tables. Each table is built from a recipe. The recipe is a JSON document that names the sources, output columns, SQL and checks.

When a run succeeds, it writes a new immutable version as one or more Parquet files. The live pointer moves to the newest version that passed its checks. Older versions remain available.

Two ways in

If someone has already published the data you need, start with the catalog. Every public dataset has a page, a documented schema, a Parquet download and a bounded JSON query API. You do not need to install anything.

• Use a public dataset
• Public Dataset API
• Connect an AI tool over MCP
• Browse the catalog

To build your own dataset, install mr-data, sign in, register a recipe and run it. The CLI sends the work to the hosted backend. Nothing runs on your machine. The package also installs the mr-data-build agent skill, which Claude Code or Codex can use to write the recipe.

• Install and sign in
• Build your first dataset
• How a build works

The sections

Section	What is in it
Start	Install, sign in, build one dataset end to end, and the objects a build is made of.
Build a dataset	Probe a source, write a recipe, run and inspect, publish and keep it fresh, source credentials, live streams, many-page sources, documents and files, fix a failed run.
Recipes	Complete recipe documents to copy: one CSV file, a JSON API, many HTML pages, PDFs, weather GRIB2, a WebSocket venue, a table-driven stream roster, two sources joined, daily city temperatures, market settlements joined, a forecast checked against observations, a replaced snapshot, many weather stations, stream trades aggregated into bars.
Use datasets	Read a published dataset from a browser, a script, or an AI tool.
Reference	Every CLI command, the recipe document field by field, sources, readers, transforms, checks, units, run states and error codes, limits.

If you are an agent

Start with For agents. It explains the command protocol, the JSON object printed by every command under --json, and the error codes an agent can branch on.

---