AI Agent Grounding

An agent that acts on infrastructure data has a failure mode humans catch by instinct: the empty enrichment result. A flat lookup API returns the same nothing for a host no feed has ever seen and for a host that is positively known clean, and a language model will fill that gap by guessing. WhisperGraph makes coverage explicit. An empty verdict means the indicator is not listed at that granularity; it does not certify the host as safe. An agent that reads coverage can escalate or hedge instead of asserting clean.

Grounding means the agent queries the graph itself instead of relying on what it remembers. Connected over MCP, it pulls a coverage-qualified verdict, the evidence behind it (explain() returns the factors and the exact feeds, each with weight and first/last-seen timestamps), and the structure around a host: hosting, routing, ownership. Every claim in its answer traces to a sourced edge. The same footing lets an agent pivot from a MITRE ATT&CK technique into the real hosting behind it; see Actor Attribution & ATT&CK.

Run it in your browser

Ground your AI agent before it acts demonstrates the grounding contract in one run: a coverage-qualified verdict band for a host, vendor identification when the host is known, and the structural context around it. The rule it teaches an agent: treat no-data as unresolved and gate on coverage. It opens with a result already loaded; anonymous runs are capped at 2 hops, and a free API key raises that to 3, no card required.

Set up your agent

Setup lives in the AI section:

• AI & Agents covers what the MCP server offers and why to ground agents in the graph.
• MCP Setup connects Claude, Cursor, or any other MCP client.
• Agents that speak plain HTTP can call the Cypher API directly with an X-API-Key header.