Skip to content

Connect your tools

The Streamlit UI, CLI, and MCP server call the same application services and query the same stored analysis. Deterministic investigation works without a provider. Optional AI uses configured profiles and explicit privacy permissions.

Choose an interface

Interface Best for Reference
CLI Shell automation, JSON queries, controlled imports/analysis, local administration CLI guide
MCP stdio A local AI client querying a dedicated data home MCP setup
MCP Streamable HTTP Private clients with bearer authentication and workspace scope MCP transport and permissions
Provider profiles Optional source explanation, diagnostic assistance, and grounded lineage claims LLM providers

There is no standalone REST API. MCP's HTTP transport exposes MCP, not an unrestricted web API. The capability matrix records implemented routes and explicit policy restrictions, including read-only MCP maintenance and UI-only cleanup.

Query before reading source

uv run gravetrail --help
uv run gravetrail --home ./automation-data --json workspace list

With an analyzed home, resolve the workspace and asset IDs first. Check freshness, completeness and cursors, then request bounded graph/column/report results. Retrieve source evidence only when needed. Start from the agent workflows and response budget guidance for complete examples.

One process owns each DuckDB home. Stop the UI before a separate CLI/MCP process opens that home, or give automation its own home. Pass an explicit --home consistently. Read-only is the launch default; write, configuration, network and application grants are independent. An agent cannot grant itself broader access.

Configure AI deliberately

Use Connections & AI → Providers & routing to create profiles and assign tasks. Use masked session credentials or named read-only launch variables. Profile persistence stores references, never secret values. Clear session credentials when finished.

Remote requests need privacy permission; full source disclosure has its own control. Output must validate against known files, quotes and line ranges before explicit merge. Model confidence alone never makes a dependency verified.