MCP

rowlab is an MCP server: the same data work as the API, as tools, with instructions that teach the workflow. It is at /mcp on a running rowlab, or rowlab mcp over stdio.

Connect

A local rowlab (uvx rowlab) needs no sign-in:

claude mcp add --transport http rowlab http://localhost:8000/mcp    # Claude Code
codex mcp add rowlab --url http://localhost:8000/mcp                 # Codex

Cursor and other clients that take a JSON config:

{ "mcpServers": { "rowlab": { "url": "http://localhost:8000/mcp" } } }

A client that wants a command rather than a URL gets uvx rowlab mcp, which runs its own rowlab on the same data folder.

A hosted rowlab

Hosted, the client signs in with WorkOS when it first connects, so a person’s agent works as them, in their organisation:

claude mcp add --transport http rowlab https://app.rowlab.dev/mcp && claude mcp login rowlab

Or it sends an API key from Settings → API keys, for a script or a server that can’t sign in:

claude mcp add --transport http rowlab https://app.rowlab.dev/mcp --header "Authorization: Bearer rl_…"

In a JSON config, that is "headers": {"Authorization": "Bearer rl_…"} next to the URL. Someone connecting over MCP must have signed in to rowlab in a browser once, which is where they join their organisation.

Tools

Tool	What it does
datasets_list, dataset_get	the datasets, and one dataset with its columns, lineage and profile
dataset_add, dataset_delete	add a file from a path, URL or S3; delete
sql_query	rows from one DuckDB SELECT over data
columns_compare	agreement between answer columns, and accuracy against a reference
run_preview	the estimate and the rendered prompts for a run, before it costs anything
run_start, run_get, run_resume, run_stop	runs, with run_get and run_resume waiting up to 30 seconds for progress
dataset_export	a dataset as CSV, JSONL or Parquet, to a path or S3
task_types_get, models_list	the answer types and their fields; models by name or kind

Each tool’s description says why it exists and when to use it, and the server’s instructions give the order of work: read the dataset, run on a few rows, read the answers, then run on all. The skills go further.