Loading ToolBench

Reading the current corpus…

OCR MCP server · C · ToolBench

OCR

Self-hosted document intelligence, table extraction, and OCR engine with ONNX Runtime acceleration. Provides native MCP server for AI agents, LangChain/LlamaIndex loaders, and REST API for PDF/PPTX/image processing.

2026-07-28+STDIO4 tools

62/100

Definition Quality67

Protocol Readiness50

SupportabilityNot measured-

Spec posture

Inferred effective spec 2026-07-28+·current baseline 2026-07-28

Static source inference · medium confidence

Current-spec patterns detected

Summary

Four tools with complete input schemas and descriptions, but quality is uneven. Tool names follow verb_noun pattern (blast_ocr_process, blast_ocr_extract_*) but descriptions vary in clarity and actionability. All tools have documented parameters with types, but descriptions lack LLM-optimized guidance on when to use each tool vs. alternatives. Output schemas are not documented, LLMs cannot predict return structure. Error handling and recovery guidance are absent. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite all being read-only operations.

Tools (4)

blast_ocr_extract_formulasread onlysource verified67/100

Detect mathematical formulas and convert to standard KaTeX/LaTeX Markdown delimiters ($...$ and $$...$$).

blast_ocr_extract_tablesread onlysource verified63/100

Extract structured tables from document images or scanned pages into Markdown and HTML, scored with a built-in TEDS evaluator.

blast_ocr_processread onlysource verified70/100

Run high-throughput OCR and document intelligence extraction on a PDF, image, or PPTX file. Returns full markdown text, dual-layer PDF, and structured metadata.

blast_ocr_semantic_chunkread onlysource verified68/100

Extract document text and split into hierarchy-aware RAG chunks with TOC lineage and token bounds.

Issues (5)

CRITICALtool ↗

Output schemas not documented. LLMs cannot predict return structure (markdown text, metadata fields, table format, chunk structure). Forces agents to guess downstream field names and types.

blast_ocr_processblast_ocr_extract_tablesblast_ocr_extract_formulasblast_ocr_semantic_chunk

HIGHtool-annotations ↗

No tool annotations despite all tools being read-only. Missing readOnlyHint metadata prevents agents from understanding safety profile and retry semantics.

blast_ocr_processblast_ocr_extract_tablesblast_ocr_extract_formulasblast_ocr_semantic_chunk

HIGHtool-description ↗

Descriptions lack actionable guidance on when to use each tool. 'Extract tables' vs 'extract formulas' vs 'semantic chunk', unclear which to call for a given document type or use case.

blast_ocr_extract_tablesblast_ocr_extract_formulasblast_ocr_semantic_chunk

Recommendations

Document output schema for each tool. Example for blast_ocr_process: 'Returns {markdown: string, metadata: {pages: int, language: string, confidence: float}, dual_layer_pdf: bytes, formats: {docx?: bytes, epub?: bytes}}'.
Add readOnlyHint: true to all four tools in the tool definition to signal they are safe to retry and have no side effects.
Expand tool descriptions to include 'When to use' guidance. E.g., 'Use blast_ocr_extract_tables for documents with tabular data; use blast_ocr_semantic_chunk for RAG/chunking workflows.'
Add error recovery hints to descriptions. E.g., 'If file is unsupported, try converting to PDF first. Supported formats: PDF, PNG, JPG, PPTX.'
Specify constraints in parameter descriptions: 'source_path must be an absolute path to an existing file (PDF, PNG, JPG, or PPTX). Relative paths and non-existent files will return an error.'

Score history

Overall score trend

First recorded score · v2 rubric

62/100

Scored	Grade	Overall	Spec posture	Rubric
2026-09-23	C	62	2026-07-28+	v2