Humsana adds real-time cognitive awareness to AI tools — so risky actions don't happen at the worst possible moment.
Install in minutes
pip install humsana-daemon
npx @humsana/mcp-server
They make them when they're sleep-deprived, cognitively overloaded, context-switching, or rushing late at night. AI systems can't see this.
Humsana introduces a cognitive safety interlock between users and AI execution.
Lightweight process monitors typing rhythm, corrections, and activity — never content.
Stress, focus, and fatigue computed on your machine. No cloud required.
Claude, Cursor, or custom agents get structured context about how you're doing.
Destructive operations blocked. Overrides logged. Audit trail preserved.
Four tools available to Claude and other MCP-compatible agents:
Returns stress_level, focus_level, fatigue_level, uptime_hours, and response style recommendations.
Checks if a command matches dangerous patterns and whether user fatigue exceeds threshold.
Executes shell commands with interlock protection. Blocks when fatigued, logs overrides.
Blocks large destructive rewrites (>30 lines removed) when user is fatigued. Saves to review folder.
Example response from get_user_state:
{
"state": "focused",
"state_label": "🎯 Deep Focus",
"metrics": {
"stress_level": 0.30,
"focus_level": 0.72,
"cognitive_load": 0.55,
"typing_wpm": 45,
"fatigue_level": 28,
"uptime_hours": 3.2
},
"fatigue": {
"level": 28,
"category": "low",
"uptime_hours": 3.2
},
"recommendations": {
"style": "direct",
"length": "concise",
"ask_clarifying_questions": false,
"tone": "efficient, don't interrupt flow"
}
}
The interlock activates when dangerous commands meet high fatigue:
rm -rf and sudo rm
DROP TABLE and DROP DATABASE
git push --force
kubectl delete
terraform destroy
docker system prune
When Humsana blocks a risky action, the user can still proceed — but only by explicitly declaring intent.
Overrides are logged to ~/.humsana/audit.log for post-mortems.
Trust is not optional.