toone

Feature · Control and safety

Safe Mode

Redact forbidden topics from Claude tool output before it reaches the conversation, block output if screening fails, and check each result in an audit log.

Product guide · Safe Mode

Redact before tool output enters the conversation

Safe Mode screens command results, file reads, and web content produced through Claude tools. When the selected redaction model finds a forbidden topic, that content is replaced before it reaches the agent conversation.

Green SAFE: this thread's tool output is screened before the model.

Define forbidden topics

Add the subjects that should not appear in tool output and customize the redaction prompt when the default policy is not specific enough for the project.

Settings → Mode: switch on Safe Mode and list the forbidden topics.

Fail closed

If screening fails, Safe Mode withholds the output instead of passing it through unreviewed. Choosing a faster screening model can change speed and precision, but it does not weaken that failure behavior.

If screening fails, the agent gets a redaction marker, not the output.

Inspect the audit trail

The audit log is a file on your Mac, not a view inside Toone. It records the screening outcome for each tool call and the post-redaction text the model received, shortened for long results. In Settings → Mode, choose Reveal Audit Log to open it in Finder. Review it when validating that the configured boundaries are behaving as intended.

Settings → Mode → Audit Log: Reveal Audit Log shows the log file in Finder.

Safe Mode currently requires the Claude provider; Codex sessions cannot use this redaction hook. Safe Mode is not available in the Mac App Store edition.

Next: Live Share.