CSV Analyzer

CSV Analyzer

Each source holds one CSV. There is no live API — the file is what the user uploaded on that source. A newer upload replaces it. If nothing is stored yet, say so; never invent figures.

When several sources are attached, pass mcpbundles_connection (the source name) on every call. Read each source note before choosing one.

Exploration workflow

1. Describe — column names, DuckDB types, row count, and a few sample rows. Always do this before assuming column names or types.
2. Profile column (optional) — nulls, approximate distinct count, top values, numeric min/max when parseable.
3. Query SQL — aggregates, filters, GROUP BY on table data (one row per CSV line).
4. Query rows — paginated raw strings when you need exact cell text without SQL.
5. Status — source name, current filename, and upload time.

SQL (query_sql)

• Table name is always data.
• Column names match the CSV header (case-sensitive); get them from describe.
• Types may be VARCHAR — use CAST(your_column AS DOUBLE) or BIGINT for math.
• Results cap at 500 rows; add LIMIT in SQL for top-N.

Response shape

• source — filename, captured_at.
• Describe — columns, sample_rows, total_rows, next_steps.
• Profile column — null_count, approx_distinct, top_values, optional numeric_range.
• Query SQL — columns, rows, returned, truncated, optional cache_hit.
• Query rows — headers, rows, offset, limit, has_more, optional total_rows.

Gotchas

• Read-only except delete — uploads happen on the source in MCPBundles.
• One file — uploading again replaces the file. Week-over-week questions work when this file already contains those weeks.
• User agent context on the source explains what the file is — read it when present.