BitssyBytes logoAgent Safety Score

Directory / MCP servers (protocol)

B

MCP servers (protocol)

Protocol · Model Context Protocol · Last checked 2026-10-16

The spec says hosts should get user consent before tool calls and treat tool descriptions as untrusted. HTTP servers use OAuth 2.1 with audience-bound tokens. Real safety depends on each host and server.

Overall: B (2.33 avg over 3 verified criteria)

Lock this down first

Only install servers you trust, and never pass a server's token through to upstream APIs.

Scorecard

CriterionRatingNotes
Permission granularity2/3
Human approval gates2/3Consent is a SHOULD in the spec. Hosts decide how it works.
SandboxingUnverifiedNot confirmed in the docs we checked.
Secrets handling3/3OAuth 2.1 with PKCE and resource indicators. Token passthrough is forbidden.
Audit logsUnverifiedNot confirmed in the docs we checked.
Kill switch / interruptUnverifiedNot confirmed in the docs we checked.
Network egress controlUnverifiedNot confirmed in the docs we checked.

Unverified: Sandboxing, Audit logs, Kill switch / interrupt, Network egress control. That means we could not confirm it in public docs, not that it is missing.

Sources