passmcp-graph¶
passmcp-graph builds a local graph of which agents use which MCP servers, which tools those servers expose, and which identities reach them, from passmcp's evidence and your MCP client configurations. It answers questions about that graph offline and gates on policy in CI.
passmcp-graph ingest attestations/ reports/ claude_desktop_config.json
passmcp-graph analyze
passmcp-graph query "agents -> tools where destructive and server.auth = none"
passmcp-graph check --policy policy.yaml
passmcp-graph export --format graphml > estate.graphml
| Command | Does |
|---|---|
ingest PATH... |
Reads files, or every *.json under a directory; see Ingesting evidence |
analyze [--format text\|json] |
Inherited risks and over-privileged identities |
query QUERY [--format text\|json] |
Paths that match; see Query language |
check --policy FILE [--format text\|json] |
Forbidden paths; see Policies |
export --format json\|graphml\|cypher |
The whole graph, to stdout |
import FILE |
Merges a JSON export into the store, after validating it |
dashboard [--out FILE] [--open] [--policy FILE] [--stale-days N] [--format html\|json] |
One offline HTML page summarising the graph; see Dashboard |
completion bash\|zsh\|fish |
Prints a shell completion script |
version |
Prints the version |
Global flags come before the command: --store DIR (default
.passmcp-graph) and --offline. Results go to stdout and diagnostics to
stderr. The exit status is 0 on success, 1 when check finds a
forbidden path, and 2 on any error.
Risk inheritance¶
A server's failing finding of severity critical, from its latest
attestation, is inherited by every agent that uses the server. analyze
reports each as the agent, the finding and the path. When a newer
attestation no longer fails the check, the risk is gone.
Over-privilege¶
An identity is over-privileged on a server when it holds a scope naming
write, admin, delete, manage or owner as a whole word (or *)
and every tool the server exposes is annotated readOnlyHint: true. A
server with no known tools flags nothing. Which tools an agent actually
calls is not recorded, so the judgement is against what the server
exposes.