In-band MCP sleeper sentinel — mid-session diff and quarantine inside your own client's session.
Scored across 2 tools
The two tools have clearly distinct purposes: sentinel_inspect runs analysis over a transcript, while sentinel_verdicts lists stored verdicts. Although inspect also returns verdicts, the action of running versus listing is evident from names and descriptions.
Both tools use a consistent snake_case sentinel_ prefix. The suffixes differ in type (inspect is a verb, verdicts is a noun), but the overall pattern remains predictable.
Two tools is quite thin for the apparent sentinel-management domain. The count is not unreasonable for a narrow inspection surface, but it feels borderline and may push agents into workarounds.
The surface covers running an inspection and listing stored verdicts, but lacks operations for retrieving a specific verdict, clearing or deleting stored verdicts, or configuring sentinel behavior. These are notable gaps for managing per-(server, client) verdicts over time.
Run the in-band sentinel over a session transcript; returns the per-(server, client) verdict and the context-safe delivered messages.
Inputs: client (required), server (required), messages (required), admission_complete.
List your stored per-(server, client) verdicts.
Inputs: no named parameters.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Diff and poisoning detection for MCP servers: catch schema drift and prompt poisoning.
Owned MCP research honeypot: synthetic decoy data with canary markers; digest-only logging.
Security research canary remote MCP server for owned-account testing.
Concurrent two-client overlap probe proving MCP servers do not cross-deliver tool results.
Add one secure layer between your agents and this server.