The server covers the full loop: create a dataset, load a local file into it, open a session, then ask questions as jobs. All 9 tools ship with an empty description, so the parameter names are the whole contract — job_mode and output_language are accepted without the server saying anywhere which values it takes. Removal is dataset-level; mcp_powerdrill_delete_dataset has no per-data-source counterpart, so a file loaded by mistake stays until the dataset goes.
A client for Powerdrill datasets: it creates them, feeds them local files as data sources, and runs question-answering jobs against them inside a session.
- Dataset lifecycle in one place — create one from a name and description, list what exists, read a single dataset's overview, delete one by id.
- Local files turned into data sources, with the file path, file name and chunk size chosen at upload time.
- Sessions that hold job history, opened with an agent, a job mode, an output language and a cap on how much history stays in context.
- Question-answering jobs that run against a dataset and named data sources and can stream the answer back.
- Paged listings with their own filters: a search term for datasets and sessions, a status filter for data sources.
A Powerdrill account reachable with POWERDRILL_USER_ID and POWERDRILL_PROJECT_API_KEY, and the files you intend to load sitting on the same machine — mcp_powerdrill_create_data_source_from_local_file reads a local path.
One command plus a key — npx -y @powerdrillai/powerdrill-mcp@latest, then supply credentials