MCP server for document processing operations using Nutrient DWS API, providing tools for PDF manipulation, OCR, text/table extraction, redaction, signing, and format conversion
This is a STDIO-only document processing skill with 16 tools. While tool names follow verb_noun conventions and descriptions are present, the evaluation reveals critical gaps in schema completeness, parameter documentation, and error handling guidance. The codebase shows basic functionality but lacks production-grade polish. Many tools have descriptions but parameter annotations are often minimal. Most critically, visible parameter schemas are incomplete or missing type information for key parameters (e.g., 'pages' object type lacks structure definition in extract_key_value_pairs, extract_table, extract_text). The server exposes file I/O operations without comprehensive validation or recovery guidance. Schema quality varies widely across tools, with some (e.g., convert, rotate) having reasonable structure but others (e.g., sign, password_protect) lacking detailed parameter documentation. No tool annotations (destructiveHint, readOnlyHint, idempotentHint) are present despite many tools performing irreversible operations.
Add blank pages to a PDF
Convert a document to another format
Stage AI redaction annotations for review; never apply them
Delete specific pages from a PDF
Create a new PDF from specific page indices (supports duplicates/reordering)
Extract key-value pairs from forms or structured documents
Extract table data from a document
Extract text from a document as JSON, with optional plain-text export
Pages parameter schema undefined across 6 tools (extract_key_value_pairs, extract_table, extract_text, create_redactions_ai, rotate, split). Parameter documented as 'object' with 'optional start/end keys' but no formal JSON Schema structure, leading to ambiguous interpretation (zero-indexed? inclusive bounds? type of values?).
Destructive operations (delete_pages, duplicate_pages, password_protect, sign) lack error handling guidance, recovery suggestions, or confirmation patterns. No guidance on what happens when operation partially fails or how to undo changes. LLM has no recovery path on failure.
No tool annotations present (readOnlyHint, destructiveHint, idempotentHint). Tools performing irreversible operations (delete_pages, split, sign, password_protect, merge) should be explicitly marked destructive so agents understand side effects and retry implications.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 43 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 37 | - | v1 |
Merge multiple documents into one PDF
Run OCR on a document
Optimize PDF size and quality
Protect a PDF with user/owner passwords
Rotate pages in a PDF
Digitally sign a PDF
Split one PDF into multiple PDFs by page ranges
Add a text watermark to a document
Output schemas not documented for any tool. Callers don't know what structure to expect. For example, extract_text's output format is undefined, is it plain text string, JSON with metadata, or structured tokens? This forces LLMs to guess and wastes tokens on discovery.
Parameter descriptions are sparse. 'pages' in split described only as 'List of inclusive page ranges in start:end format', no guidance on how to handle overlapping ranges, out-of-bounds indices, or invalid format. Parameters lack constraints (min/max, pattern, enum).
Type union in ocr language_arg ('string|array') is not formally JSON Schema. Should use oneOf or proper type constraint. Also no list of supported languages provided, forcing LLM to guess or experiment.
password_protect parameters refer to reading passwords from 'owner-only local file', a serious security anti-pattern. Credentials should use server-side secret injection via environment variables, not file paths exposed in tool parameters.
convert tool accepts format parameter as free-form string ('pdf, pdfa, pdfua, docx, xlsx, pptx, png, jpeg, jpg, webp, html, markdown') but no enum constraint in schema. LLMs commonly hallucinate unsupported formats. Should declare enum with exact supported values.
Tool descriptions lack recovery or dependency hints. E.g., 'ocr' has no explanation of when OCR is needed (scanned PDFs vs digital PDFs) or what to do if OCR fails. 'merge' doesn't explain page ordering or metadata handling.