Deterministic Edge Control for AI Agents via Model Context Protocol over MQTT

How we bridge probabilistic AI models with deterministic hardware endpoints using strict JSON schemas over MQTT for safe, reliable real-time edge computing.

The integration of AI agents with IoT represents a paradigm shift from static, rule-based automation to dynamic physical orchestration. However, bridging probabilistic Large Language Models with deterministic hardware endpoints introduces severe reliability challenges - most notably the risk of models generating structurally invalid parameters during execution. To address this, we propose a standardized integration architecture utilizing the Model Context Protocol (MCP) over MQTT. By explicitly constraining the AI action space with strict JSON schemas, the framework prevents hallucinated or malformed commands. Transmitting structured JSON-RPC payloads over MQTT ensures minimized jitter, high temporal consistency, and the deterministic low latency required for safe, reliable real-time edge computing. This research was presented at the Anniversary International Scientific Conference CTA 2026 in Pamporovo and is supported by project SP25-FMI-008 at Paisii Hilendarski University of Plovdiv.

Source: View original publication (PDF)