[Feature Request] Add Agent skill/tool integration interface for OCR capability

Motivation

Unlimited-OCR provides powerful document parsing capabilities (single image, multi-page, PDF) through Transformers, vLLM, and SGLang. However, there is currently no standardized way for Agent frameworks (e.g., LangChain, DuMate, AutoGPT, CrewAI) to integrate and invoke Unlimited-OCR as a callable skill/tool.

As LLM-based agents become increasingly popular for document processing workflows (RAG pipelines, automated document analysis, intelligent assistants), the ability to call Unlimited-OCR as a structured tool — rather than wrapping the Python API manually — would significantly broaden adoption and ease of integration.

Current State

• Python API: model.infer() / model.infer_multi() — requires importing the model directly
• vLLM deployment: OpenAI-compatible server, but OCR-specific parameters (crop_mode, ngram settings, output format) are not exposed as a standard tool interface
• SGLang deployment: similar server-based approach
• No MCP (Model Context Protocol) or function-calling tool definition is provided

Proposed Feature

Add a standardized skill/tool interface that allows Agent frameworks to invoke Unlimited-OCR as a tool. Suggested approaches (any or all):

Option A: MCP Server

Provide an MCP (Model Context Protocol) server that exposes OCR capabilities as tools:

• ocr_parse_image — parse a single image (supports gundam and base modes)
• ocr_parse_multi — parse multiple images / pages
• ocr_parse_pdf — parse a PDF document (auto-convert pages to images)

Each tool should accept standard parameters (image path/URL, output format, crop mode, etc.) and return structured results (parsed text, layout info, confidence scores).

Option B: Tool/Function Definition

Provide an OpenAI-compatible function/tool definition (JSON schema) that can be registered in Agent frameworks:

{
  "name": "unlimited_ocr",
  "description": "Parse documents (images, PDFs) using Unlimited-OCR. Supports single image, multi-page, and PDF parsing with configurable crop modes.",
  "parameters": {
    "type": "object",
    "properties": {
      "input_type": {"type": "string", "enum": ["image", "multi_image", "pdf"]},
      "input_path": {"type": "string", "description": "Path or URL to the input file(s)"},
      "crop_mode": {"type": "boolean", "default": true, "description": "Enable gundam crop mode for single images"},
      "output_format": {"type": "string", "enum": ["text", "markdown", "json"], "default": "markdown"}
    },
    "required": ["input_type", "input_path"]
  }
}

Option C: REST API Wrapper

A lightweight HTTP API (FastAPI or similar) that wraps the existing Python API with standardized endpoints:

• POST /ocr/parse — accept image/PDF input, return structured OCR output
• Support for both file upload and URL-based input
• Configurable parameters via request body

Benefits

1. Ecosystem integration — enables plug-and-play usage in LangChain, DuMate, AutoGPT, and other agent frameworks
2. Standardized interface — no need for each user to write custom wrapper code
3. Structured output — returns machine-readable results (JSON/Markdown) instead of raw model output
4. Wider adoption — lowers the barrier for agent-based document processing workflows

Additional Context

• The existing model.infer() and model.infer_multi() APIs are well-designed but tightly coupled to the Python runtime
• vLLM/SGLang server deployments already provide HTTP interfaces, but lack tool/function definitions for agent integration
• MCP is becoming the de facto standard for tool integration in AI agents

Would the maintainers be open to this? I'm happy to contribute to the implementation.