Run live no login, a few runs per hour
Custom task needs the demo access code
Tasks
How it works
- An LLM agent picks tools (
read_page,propose_title,finish) within hard step, page and token limits. The execution layer, not the model, enforces the rules. - Every page opens in a fresh Docker sandbox: no secrets, read-only disk, no capabilities, CPU, memory and time limits.
- The sandbox reaches the internet only through an egress proxy with an allow list. Everything else is blocked and logged.
- Only a text summary and the source URL come back. Instruction-like text is quarantined and never reaches the model.
- A human approves the new title before anything can be exported.