Connect your AI over MCP

ST-1 exposes its primitives as a Model Context Protocol (MCP) server. Any client that speaks MCP — Claude Desktop, Claude.ai web, Cursor, VS Code Copilot, Goose, the CLI — can connect, see your projects, and call tools on them.

Connecting

From the app, open Connected agents. The page shows the connection URL and an OAuth flow. Paste the URL into your client’s MCP server settings; the client opens a browser window to authorize, and your projects show up as callable tools.

What the AI can do

The MCP surface mirrors the web UI:

• Projects — list, create, rename, describe, delete.
• Datasets — upload, profile, sample, query, quality-report, set descriptions, delete.
• DBTs — create from SQL, update, delete.
• Visualizations — create, update, get, list, delete. Vision-capable clients receive a rendered PNG of the chart inline; capable clients (Claude.ai web, Claude Desktop, etc.) get an interactive resource that renders the chart directly in chat.

Vocabulary the AI uses

• Source — an uploaded dataset.
• DBT — a SQL-defined derived table over sources or other DBTs. Not dbt-core.
• Project — workspace containing sources, DBTs, and visualizations.

Recommended workflow

When given new data, your AI should:

1. upload_dataset to ingest.
2. get_dataset to read the column profile (nulls, distincts, ranges, hygiene flags).
3. quality_report to surface concrete issues with ready-to-paste cleanup SQL.
4. create_dbt named <source>_clean to bake the fixes — cleanup is always a DBT, never a parallel mutator API.
5. Build analysis DBTs and call create_visualizationon top.

Live updates

Mutations the AI makes fan out over server-sent events to any browser tab open on that project — so as the AI uploads, creates, and visualizes, you see it appear in your browser without a refresh.