Safety audit trail

Beta — the MCP Gateway is open source and on npm: npx -y @mcp-rating/gateway. Source →

The audit trail answers one question — "did this server try anything?" — with evidence. It's a forensic log of what every server did and what the sandbox blocked.

What it records

Event Meaning
connect A server was connected, with the capabilities it was granted, or restarted in the same sandbox after a crash
tool_call A tool ran, and whether it succeeded
tool_blocked A tool call was refused by your permission policy or a plugin
egress_blocked An outbound connection was refused, with the target host
egress_allowed An allowlisted connection was permitted
personal_data_requested A tool result appears to ask for personal data, with the kinds it matched (a heuristic)
disconnect A server was torn down

It records what was contained — not your arguments, results, or secrets. It's a security artifact, distinct from usage analytics.

Querying

From any client, use the mcp_audit meta-tool. For the security-reviewer view — only blocks and errored calls:

Use mcp_audit with securityOnly true

Durable log

Set an environment variable to append every event to a JSONL file you can ship to a SIEM:

MCP_GATEWAY_AUDIT_LOG=/var/log/mcp-audit.jsonl npx @mcp-rating/gateway

See also