Skip to content

Behaviour to be aware of

From docs/MCP_TOOLS.md on GitHub

These are properties of the current code (0.6.0) that callers should know about; each was checked against the capture run and is described in more detail in the entry named.

  • The relationship graph keeps one edge per pair of nodes. Parallel security group, NACL and internet-exposure rules are merged into one edge with comma lists in port_range and protocol, so graph_stats.graph_edges, subgraph_export and the graph resources show fewer edges than get_edges, which reads the raw list (Graph).
  • "Internet reachable" in top_risks, reachability_findings, internet_exposure(include_reachable=true) and the attack paths follows network-flow edges only, and the path tools add the identity pivots explicitly. internet_exposure items and find_assets(internet_exposed=true) use the collector's flag only.
  • Error messages never contain the value the caller passed or names from the dataset; read cloudg/error_data (value, suggestions, valid) for them (Errors).
  • sparql_query, subgraph_export and the text graph and ontology resources are restricted, so any profile with a confidential ceiling (strict, the soc-analyst analyst) hides them; cloudg://graph/d3 stays available.
  • security_coverage checks vulnerability scanning for EC2, container registries and Lambda only.
  • cross_account_edges identifies endpoints by ARN where other graph tools use ids; both forms are valid references.
  • The live guard's cooldown applies per scope to every caller, so a second agent asking for the same collection within the cooldown gets -31029 and the name of the dataset to use instead.