If you live in Cursor (or Claude Code, or Codex) all day, the IDE feels like the control plane.
It isn't — not once you walk away.
Decision 1: Can this tool call run?
Agents with shell access don't fail like unit tests. They fail like ops incidents: the command already ran.
A useful outer loop is boring:
- Agent proposes a tool call
- Human gets approve / deny before execution
- Deny means stop — not "log it for later"
I put that path on a free web dashboard: pair a Mac once (connector dials out over HTTPS — no inbound ports), open a browser, approve or deny. That's Web Control at $0/month on ThumbGate.app while the machine is online. Leash-style approvals ride that free path — not a fake "pay to stop the agent" upsell.
Decision 2: If the Mac disappears, who finishes the work?
Chat UIs skip this. Real operators need an offline policy:
- Pause
- Ask
- Continue on a fenced VPS runner (paid Continuity)
I ship Continuity, and I will not pretend it's fully battle-proven. Live product copy still says we're proving it out in real use. Price is Live price in-product (Stripe-driven) — don't trust a blog post for a dollar figure.
Team tier includes a Priority Fly.io cloud VPS runner as a listed feature. ThumbGate is not affiliated with Fly.io (or Cursor, Anthropic, OpenAI, Nous Research). Naming the tech is not a partnership claim.
Adjacent work: credentials vs. tool approval
Last week's HN attention on credential gateways for agents is healthy. Keeping secrets out of prompts is necessary.
It is not sufficient. Correct credentials on a bad command still hurt.
If your pain is runaway loops and pre-action gates rather than remote control, start here:
Also on phone
Same remote for Hermes agents, paid store download (verify live price on the store page):
Building in public
Almost no real customers yet. If "two decisions" is the wrong product shape, tell me — that's more useful than a polite like.
Try the free browser path: thumbgate.app
Top comments (0)