Run your first
Agentic OS.
See a task pass through planning, drafting, verification, rules, and human approval. This working browser sandbox shows the control flow with realistic failure handling—without connecting to a model or sending your notes anywhere.
Use invented notes only. This demonstration is not Claude, not a production agent, and not a security sandbox. Source excerpts can look plausible even when facts are wrong; check them yourself.

Choose a practice task
Each line represents one untrusted source note. Use the good example first, then try the hostile example and see which gate stops the result.
Rulebook: read-only input, at most six notes, exact-quote checks, no external actions, no auto-approval.
Inspect all six gates
No run yet.
Approval here is a local practice decision by the learner. It never sends or publishes a message.
Proposed internal draft
The local draft appears after a practice run.
Trace claims to sources
Original quotations and verifier checks appear here.
Ready for a real Claude-generated draft?
Download the optional Python runner and follow the nine-step setup guide. It uses your private API key on your own machine, requires explicit paid-call confirmation, and keeps all generated output draft-only until a human reviews it. The website never collects your API key.
python3 agentic_os_claude.py --state ./live-state plan --task demo-safe.json
python3 agentic_os_claude.py --state ./live-state run --task demo-safe.json --confirm-paid-api-use
python3 agentic_os_claude.py --state ./live-state reportSupervised pilot only: no authenticated external connectors, automated posting, remote execution or production autonomy are provided by this teaching kit.