netsuite-saved-search-mcp

NetSuite saved search exports look like Excel files but are actually XML SpreadsheetML with a .xls extension — Excel opens them, pandas trips on them, and most teams writing Claude integrations against NetSuite end up rebuilding the same parser by hand. This MCP server exposes those exports as a structured query interface for Claude: discover files, inspect columns, filter and aggregate rows, categorize by memo keywords, surface anomalies. Seven tools wrap a parser that handles the format's real quirks — entity-encoded text, empty cells flagged with ss:Index, drifting column layouts between saved searches, DateTime cells that won't parse. The parser runs in lxml's recovery mode and reports broken cells as structured warnings rather than crashing on them.

What an auditable agent run looks like

A condensed exchange reproduced from the claude-eval-harness NetSuite suite's anomaly_zero_activity_june case (runs/baseline-sonnet-4-6.json). The agent's tool calls are auditable, the parser's total_supporting_count is citable, and the model never has to fabricate numbers it can't point at.

user      What stands out in this GL? Run detect_anomalies on
          sample_gl_export.xls (account_column=Account,
          amount_column=Amount, period_column=Period).

agent     → detect_anomalies(file_path="sample_gl_export.xls",
                             account_column="Account",
                             amount_column="Amount",
                             period_column="Period")
          ← {findings: [
              {severity:"HIGH", category:"zero_activity_period",
               description:"No rows recorded for Jun 2024",
               total_supporting_count:0},
              {severity:"MEDIUM", category:"ratio_anomaly",
               description:"Account 1200 in Sep 2024 totals 319,219.74
                            — 3.1x the 1200 median of 103,259.01",
               supporting_rows:[...10 rows...],
               total_supporting_count:17}]}

agent     HIGH: no rows posted to June 2024 inside an otherwise continuous
          range — likely an extraction gap or skipped close. MEDIUM:
          Account 1200 spiked to 3.1× its own median in September across
          17 cited rows. supporting_rows trims to 10; total_supporting_count
          carries the full count so it's verifiable.

The full anomaly response also lives in examples/walkthrough.md.

Quick start

uvx netsuite-saved-search-mcp           # or: pip install netsuite-saved-search-mcp
export NSMCP_ROOT=/path/to/your/exports

Add to Claude Desktop's config (full version in examples/claude_desktop_config.json):

{
  "mcpServers": {
    "netsuite-saved-search": {
      "command": "uvx",
      "args": ["netsuite-saved-search-mcp"],
      "env": {"NSMCP_ROOT": "/path/to/your/exports"}
    }
  }
}

Then any tool call lands directly:

{
  "tool": "query_export",
  "arguments": {
    "file_path": "Q3_GL.xls",
    "filters": [{"op": "eq", "column": "Account", "value": "4000"}]
  }
}

Standalone Python usage

The parser is usable directly without the MCP transport — useful for notebooks, batch scripts, or pytest fixtures that don't want stdio in the loop:

from netsuite_saved_search_mcp.parser import NetSuiteExport

export = NetSuiteExport("tests/fixtures/sample_gl_export.xls")
print(export.headers)            # ['Order Type', 'Date', 'Period', ...]
print(len(export.rows))          # 212
print(export.rows[0]["Account"]) # '4000'

NetSuiteExport(path) parses the file end-to-end on construction. .rows is a list of dicts keyed by header name; cell values are typed (str, int, float, datetime.date, bool, or None). .parse_warnings exposes any recoverable issues lxml hit on the way through.

Security boundary

For audit and accounting use, the server enforces a tight blast radius on what it can touch. (Reporting channel for vulnerabilities is in SECURITY.md.)

• All reads constrained under NSMCP_ROOT. Every tool resolves its file-path argument relative to this env var (or os.getcwd() if unset). Paths that resolve outside the root raise PathTraversalError before any I/O.
• Symlink-escape blocked via realpath comparison. _resolve_under_root calls Path.resolve() on the candidate then checks relative_to(root). resolve() collapses .. segments and follows symlinks to their real target, so a symlink inside the root that points outside it fails the check.
• No writes. The parser opens .xls files for reading only. No tool writes to the filesystem.
• No network calls. Runtime dependencies are mcp, lxml, python-dateutil, pydantic — none of them dial out during a tool call.
• Stderr logs are scoped. Startup logs the NSMCP_ROOT path. Tool calls log nothing by default; row data, column values, and financial figures never reach the log handler.

Why this exists

NetSuite saved search exports use XML SpreadsheetML, not Excel binary, despite the .xls extension. Column layouts drift between saved searches, so code that hardcodes column letters breaks on the next export. Empty cells are silently omitted from each row with ss:Index attributes marking where they were, which trips naive sequential parsers. Every finance team using Claude with NetSuite ends up rebuilding the same parser. This server solves it once.

Tools

Tool	Description	Key parameters
list_exports	Scan a directory for .xls files; return one summary per file with row counts, header counts, warning counts, and detected date range.	directory
get_headers	Return column headers, their spreadsheet column letters, and the 0-indexed header row.	file_path
query_export	Filter rows by a list of predicates (AND-combined), optionally project to a subset of columns, cap results.	file_path, filters, columns?, limit?
aggregate_export	Group rows by one or more columns; compute sum/count/avg/min/max per group.	file_path, group_by, measures
categorize_by_memo	Tag every row with a _category derived from case-insensitive keyword rules across one or more memo columns.	file_path, memo_columns, rules
detect_anomalies	Three checks: zero-activity periods (HIGH), ratio anomalies (MEDIUM), document-count variance (MEDIUM).	file_path, account_column, amount_column, period_column
get_parse_warnings	Return parse warnings (phantom_column, bad_datetime, encoding_recovery, empty_row_skipped) captured during parsing of the specified file.	file_path

Predicates are a discriminated union keyed on op. Example query with two predicates:

{
  "file_path": "deferred_commissions_2024.xls",
  "filters": [
    {"op": "eq", "column": "Account", "value": "1321"},
    {"op": "date_range", "column": "Date", "start": "2024-01-01", "end": "2024-12-31"}
  ],
  "columns": ["Date", "Document Number", "Amount", "Memo (line)"],
  "limit": 100
}

Measures for aggregate_export:

[{"column": "Amount", "op": "sum", "alias": "total"}, {"column": "Document Number", "op": "count"}]

Example walkthrough

See examples/walkthrough.md for an end-to-end example using the included sanitized fixtures.

Limitations

• Only handles saved search exports, not raw transaction-level XML from SuiteScript or RESTlets.
• All-string exports with no typed columns may misidentify the header row; an explicit header_row override is planned.
• Memo categorization uses case-insensitive substring matching against US-English keywords. No stemming, no fuzzy matching.
• detect_anomalies only recognises period labels in three formats — Jan 2024, January 2024, 2024-01. Quarter labels (Q1 2024) and fiscal-period labels are silently skipped. The ratio and document-count checks also need ≥3 distinct periods to produce a finding.
• Not optimized for exports larger than 100k rows. The cache holds parsed NetSuiteExport instances in memory keyed by (path, mtime) and is unbounded — a long-running session against a large directory will keep every parsed export resident.
• Coverage tested against three synthesized fixtures that exercise typed cells, lxml recovery, and 14-row-metadata header detection. Real-world saved searches with column layouts beyond those shapes may surface gaps.
• v0.1. The MCP tool schemas and the parser's Predicate/Measure models may change before v1.0.

Contributing

Issues and PRs welcome. Run uv run pytest, uv run mypy src, and uv run ruff check src tests before submitting; all three should be clean. Commits follow Conventional Commits. New tools require a Pydantic response model, a happy-path test against the included GL fixture, a failure-path test (missing file or unknown column), and an entry in the table above.

License

MIT. See LICENSE.