Data Connectors

An agent reads through retrievers and document-returning tools: a vector store behind a RAG pipeline, a search tool, an internal repository, a SaaS system exposed as a tool. Retrieval policy has authority over exactly the surfaces connected to it.

Connectors close that gap. Each one adds a read surface to what policy can see, so evaluation decides what an agent gets to know regardless of how the source itself was locked down.

Framework plugins instrument the agent's own tool and retrieval calls where the agent is built. Discovery connectors find the agents, frameworks, and MCP servers already running on your fleet. Approval channels carry a held action to a human, and export connectors stream decision records into the systems your security team already watches.

The categories matter more than any vendor list, because your agents' tools define the read path. A source an agent reaches through a tool call is a surface the platform can see.

When a retriever returns, each document is evaluated individually, content and metadata, before the model sees any of it. Evaluation runs in-process against the same locally cached rule bundle that governs actions, so the read path pays no network round trip.

Each document resolves to allow, redact, or deny. Redaction replaces protected values and lets the structure through. A denied document is suppressed without an error, a fully denied result set simply comes back empty, and genuinely exceptional access can route to a recorded human approval.

Plenty of real tools read and write in the same call: search-and-update, fetch-and-send. VisIQ treats that as one event tagged with both operations, evaluated once against every rule that applies to either.

The connector model and the rules model share one vocabulary. Sources carry trust tiers, agents carry business functions, and one engine decides across the whole surface.