What problem does it solve? Operators of a deployed NVIDIA VSS stack need a reliable, read-only way to answer analytics questions — recent incidents, VLM-confirmed violations, sensor inventories, occupancy and speed metrics — without hand-writing JSON-RPC calls against the VA-MCP server on port 9901. ## Core Features & Use Cases - Two-step MCP session workflow: Initialize a session, capture the mcp-session-id response header, then issue tools/call requests against VA-MCP tools such as video_analytics__get_incidents, get_sensor_ids, get_fov_histogram, and analyze. - Deployment gating and retry guidance: Probe VA-MCP liveness before querying, hand off to vss-deploy-profile when the alerts profile is down, and retry transient 5xx errors with exponential backoff. - Scope guard against misuse: Explicitly refuses live VLM Q&A, narrative report generation, archive search, and prompt-injection attempts embedded in untrusted alert payloads. - Use Case: Ask "How many incidents have been confirmed by the VLM verifier so far?" and the agent filters get_incidents with vlm_verdict=confirmed and reports a concrete count from Elasticsearch. ## Quick Start Ask the agent to list the sensors currently known to the video analytics stack or show the ten most recent incidents across all sensors.
