That is the detail that makes Athena workable from a conversation: a query that outruns the timeout returns its execution id, and the assistant polls the status and collects the rows later instead of losing the work. The response also carries data scanned in bytes, so you can see what a query cost before running a wider one.
A small Athena client over MCP: run a query, poll it, fetch its results, and browse the Glue-backed schema behind it.
- run_query — execute SQL against a named database with a max_rows cap; on timeout it returns the query execution id instead of failing
- get_status — the current state of a query: QUEUED, RUNNING, SUCCEEDED, FAILED or CANCELLED
- get_result — fetch the rows for a completed query by its execution id
- list_tables — every table in a database
- describe_table — column names, types and comments, plus index and key detail
- Results carry the execution metadata Athena charges on: data scanned in bytes, execution time, and the S3 output location
An S3 path for query results in ATHENA_S3_OUTPUT_LOCATION — the one required setting. AWS_REGION, ATHENA_WORKGROUP and ATHENA_TIMEOUT_SECONDS are optional. AWS credentials come from the standard chain: environment variables, an AWS CLI profile, or an instance role. Install from PyPI as aws-athena-mcp and run it with uvx. The README lists the minimum Athena, S3 and Glue permissions.
One command plus a key — uv tool install aws-athena-mcp, then supply credentials