shadow AI
AI Discovery
Agents arrive the way shadow SaaS did: an engineer wires a framework into a service or adds an MCP server to an editor config, and nobody registers either one. Discovery scans the fleet for all of it and reports what is governed, what is not, and what it could not see.
What the sensor finds
The visiq-discovery sensor reads installed packages and running services on an endpoint, passively, without modifying anything. It detects agent frameworks such as LangChain, LlamaIndex, and CrewAI through npm and pip packages, along with MCP servers, local model runtimes like Ollama, and tool-routing services.
It also walks code projects. A project that declares an agent framework gets a governance badge: ungoverned, harnessed, or governed, where governed means the VisIQ harness is installed and a registered agent on that host has called home within 7 days. A project that only declares the harness stays below governed until an agent actually reports in.
Two ways onto a device
There is one payload and two delivery shapes. A Deployed Sensor is the binary installed persistently as a managed application through your device-management platform's app catalog: it stays resident, self-heals, and scans on a schedule you set.
An On-Demand Scan delivers the same binary once, through the management platform's own execution channel. It fetches, runs, reports, and leaves nothing installed. Both paths run the identical binary, so results are comparable across the fleet, and the delivery script verifies the binary's SHA-256 before running it.
"No findings" has to be earned
A clean result is meaningless if the scan could not see clearly. Scans cut short by a time or file budget are reported as incomplete, and an incomplete scan never auto-resolves an open finding. Failing to see a problem is different from confirming its absence.
The privacy contract is narrow on purpose. The sensor reports project names, the names and versions of known AI-framework dependencies, and the VisIQ harness packages. No other dependency name and no file content leaves the endpoint.
Inventory is the precondition for policy
You cannot scope an agent to its business function before you know the agent exists. Every governance decision downstream, from credential scoping to runtime rules, assumes an inventory built from what is actually installed and running.
Each finding closes one of three ways: register it, so it is governed from then on; acknowledge it as known and intentional; or flag it for investigation. The open count is a number your program can be measured by.