list_labels and describe_label are the ones to call before writing Cypher: they return the labels that exist with their counts and the property keys actually observed on one, which is what keeps a query from filtering on a field the graph has never had. Past that, run_recipe usually beats hand-written Cypher — 29 curated recipes cover the common investigations, and the direct half of them run rate-limited with no key at all, so you can try the graph before signing up. If you self-host over HTTP, remember that the server checks nothing inbound and forwards whichever key the caller presents, so it needs your own gateway in front rather than an open port.
An MCP server over WhisperGraph, a Cypher-queryable graph of internet infrastructure: 7.39B nodes and 39B edges spanning domains, hostnames, IPs, ASNs, prefixes, organizations and web links, with 5.6M threat-intel edges on top. All eight tools are read-only, and a catalog of curated recipes means most investigations need no hand-written Cypher at all.
- query runs Cypher against the graph, validated against a safety rule set before it reaches the backend
- list_labels and describe_label are the pair that keeps a query honest: every label with its count, then one label's actual property keys — which is how you learn there is no DOMAIN or FQDN label, only HOSTNAME
- explain_indicator takes an IPv4, IPv6, hostname, CIDR or ASN, works out which it is, and returns a threat score with a level from NONE through CRITICAL plus the factors and sources behind it
- whisper_history returns historical WHOIS or BGP for an indicator — for a prefix, the origin, the start and end times and the peers that saw it
- domain_variants runs 14 mutation algorithms — omission, repetition, transposition, QWERTY-adjacent replacement and insertion, vowel swap, bitsquatting, homoglyph confusables and more — then checks which of those variants actually exist in the graph
- list_recipes and run_recipe open the catalog of 29 curated recipes: keyless direct procedures including assess, identify, explain, variants, origins for CDN de-cloaking, history and walk, plus keyed multi-step flows like attack-path, attack-surface, subdomain-takeover, bgp-hijack-exposure, blast-radius and build-takedown-evidence-package
- Six resources attach without spending a call — the full schema, the relationship map, a Cypher function reference and a query cookbook, plus live whisper://stats and whisper://quota
- Eight prompt templates cover the standard investigations: investigate-ip, map-attack-surface, compare-domains, blast-radius, threat-triage, whois-pivot, bgp-investigation and typosquat-sweep
Node 20 or newer. Run over stdio as npx -y @whisper-security/whisper-graph-mcp with WHISPER_API_KEY, or point a client at the hosted server at https://mcp.whisper.security with an Authorization: Bearer header. A key can be obtained without a browser: POST to console.whisper.security/api/signup, then POST the emailed code to /api/signup/verify, and the response carries api_key, mcp_url, dashboard_url and docs_url. There is a free trial and paid tiers for larger quotas. Self-hosting runs the same image with MCP_TRANSPORT=http on port 8080 — in that mode the server authenticates nothing inbound, relaying the caller's X-API-Key or Authorization header upstream and falling back to WHISPER_API_KEY, so it belongs behind your own gateway, with WHISPER_ALLOWED_HOSTS set as the Host allowlist. Apache-2.0.
One command plus a key — npx -y @whisper-security/whisper-graph-mcp, then supply credentials