Grasshopper

Autonomous Browser Agent with Safety Gate & BudgetLedger

60-Second Jury Replay

📺 Step-by-Step Replay Viewer

Inspect actual browser actions captured during live runs without running a browser.

Step frame
Step 1 of 6

🛡️ Trust & Safety Architecture

  • Approval Gate: Sensitive actions (buy, post, export) require human sign-off via voice, Telegram, or Web UI.
  • BudgetLedger: Hard daily spending cap ($0.50) and per-run limits ($0.05). Refuses calls when exceeded.
  • Domain Allowlist: Strict CODE_ALLOWLIST blocks unauthorized or dangerous hosts.
  • Blast Radius: Every task outputs audited file count, domain list, and cost in blast_radius.json.

⚡ Live Benchmark Summary (N=3)

ScenarioGoal1st Run LLMCost
R1Books to scrape (search & wiki)22 calls$0.0038
R2Saucedemo e-commerce flow3 calls$0.0001
R3Hacker News top 3 digest2 calls$0.0004
R4arXiv API search query6 calls$0.0006
R5The-Internet resilience suite5 calls$0.0001