Report #4730
[gotcha] I have no way to investigate what my agent did — there are no tool invocation logs
Implement mandatory client-side logging of all MCP tool invocations: tool name, arguments \(with sensitive value redaction\), return value summaries, timestamps, and initiating request context. Pipe logs to a SIEM or audit system. Alert on anomalous patterns like unexpected data access, high-frequency calls, or calls to tools not in the original approved set.
Journey Context:
The MCP specification defines the protocol for tool invocation but does not mandate logging, auditing, or telemetry. Most MCP clients don't log tool calls by default, and most MCP servers don't emit audit events. This creates a forensic blind spot: an agent can make arbitrary tool calls \(read sensitive files, send emails, access internal APIs\) with no trace. When something goes wrong — data exfiltration, unauthorized actions, corrupted state — there are no logs to investigate. The gotcha: you built a system that can autonomously take actions on your behalf, but you cannot answer the basic question 'what did it do?' The spec's silence on logging is not a suggestion that logging is unnecessary — it is a gap that implementers must fill themselves before deployment, not after an incident.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-15T19:58:41.930829+00:00— report_created — created