TableTidy — Messy CSV to Clean Typed Data (MCP)

Hand it an ugly spreadsheet, get back clean typed rows an agent can actually reason over.

TableTidy is an MCP server for AI agents. Real-world CSVs have title rows above the header, mixed types in a column, junk/separator rows, and inconsistent booleans and dates — enough to choke an agent trying to use them. TableTidy fixes that in one call.

Tools

• clean_table(csv) — the full clean. Detects the real header (skipping preamble/title rows), normalizes headers to snake_case, infers a per-column type, coerces values, and drops junk rows. Returns:

  { "schema": [{ "name": "age", "original": "Age", "type": "integer" }],
    "rows": [{ "age": 36, "active": true }],
    "headerRowIndex": 1, "droppedRows": 2, "truncated": false, "notes": [] }
  

• infer_schema(csv) — schema only (header + column names + types), no rows.
• coerce_types(csv, schema) — apply a schema you supply (column → type) instead of inferring.

Types

integer, number, boolean, date, string — inferred by sampling each column. Booleans accept true/false/yes/no/y/n/t/f; numbers tolerate thousands separators; dates normalize to YYYY-MM-DD. Blanks become null.

Notes & limits

• Stateless — nothing is stored between calls.
• Pass CSV/TSV as inline text. Rows are capped at 10,000 per call (flagged in notes when truncated).
• Header detection and junk-row removal are heuristic. They handle the common mess well; pathologically ambiguous layouts (several plausible header rows) may need a hint.