Product guide · Safe Mode
Start here
OverviewGetting startedGetting started
Install and connectCreate a projectCreate an agentCreate a routineRun and debugUnderstand Toone
Core conceptsFeatures
OverviewExperience
Agent SpotlightWorkspace windowsZen modeProject switcherProject ExplorerWorkflows
Browser sessionsRoutinesAgent orchestrationWorkflow libraryCalendarBackground threadsControl and safety
Project HistorySafe ModeConnect
Live ShareVoice and meetingsMobile companionMCPs and integrationsAI providersRedact 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.
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.
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.
Inspect the audit trail
The audit log records the screening outcome for each tool call and the post-redaction text the model received. Review it when validating that the configured boundaries are behaving as intended.
Safe Mode currently requires the Claude provider; Codex sessions cannot use this redaction hook.
Next: Live Share.