Year: 2026
Type: article
Abstract: Contemporary assistive technology for users with visual impairments relies on platform-specific screen readers thatcommunicate with applications through fragmented, vendor-defined accessibility APIs. This architectural fragmentation prevents AI agents from navigating heterogeneous digital environments in a consistent, reliable manner. This paper presents MCP-ARIA, a standardized server architecture that leverages the Model Context Protocol (MCP) to expose ARIA-compliant accessibility trees as structured tool outputs consumable by large language model (LLM) based agents. The proposed framework introduces three primary contributions: (1) a canonical JSON-LD schema that maps Web Content Accessibility Guidelines (WCAG) roles, labels, and focus sequences across web, desktop, and mobile platforms; (2) a latency-optimised request-response pipeline that reduces round-trip navigation overhead to below 120 ms compared to 1,840 ms for screenshot-based alternatives; and (3) persistent MCP resource primitives that encode user disability profiles to enable session-aware, personalised agent behaviour. Evaluation over 2,400 structured navigation tasks drawn from four heterogeneous environments yields an average task-completion accuracy of 91.7%, a macro-averaged F1 score of 0.91, and a mean accessibility-role classification precision of 0.92. These results demonstrate that a unified open protocol cansupplant fragmented native APIs while simultaneously satisfying real-time latency constraints required in interactive assistive contexts.
Language: English
Subfield: Human Factors and Ergonomics
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7168090328