Summary
Bounded output is the feature, not a limitation.
A SELECT against a real warehouse can return more than any conversation can hold; the row, byte and cell limits mean a careless query fails usefully instead of burning the context window. Job runs are covered end to end, including reading the output of a run that failed.
What it is
A Python server for Databricks that puts SQL Warehouses and the Jobs API in front of an agent, with output limits so a wide result set does not flood the conversation. Multiple workspaces are configured at once and selected per call.
What you get
- execute_sql_query with configurable row, byte and cell limits, returned as markdown tables
- discover_schemas, discover_tables, describe_table and get_table_sample for finding your way around
- connection_health to check the workspace is reachable before anything else
- list_jobs, get_job_details, get_job_runs, trigger_job, cancel_job_run and get_job_run_output
- cache_stats and performance_stats — hit rates, operation latencies and error rates
- Retries and circuit breakers around the API calls
- Workspace switching by one parameter, so dev and prod are the same conversation
Requirements
Python 3.10 or later, installed with uv sync or pip install -e .. Workspaces are declared in an auth.yaml file — each with a host, a token and an HTTP path to a SQL warehouse — with one marked as the default.