Guide · agent operations
Monitor Claude Code from your phone
Background coding agents do hours of real work — and then stall at a permission prompt on a desk you already left. Here is what remote monitoring actually requires, what people use today, and what a purpose-built terminal changes.
The problem: agents work, then wait
Claude Code, Cursor background agents, Devin-style autonomous coders — the pattern is the same. You delegate a task, the agent runs the loop (reproduce, patch, test, verify), and then it hits something consequential: a production deploy, a force-push, a migration. It stops and asks. If you are at lunch, in a meeting, or asleep, the work sits finished-but-unshipped for hours.
The bottleneck is not the agent — it is the approval living on one machine. Monitoring from a phone is really three separate needs: seeing live progress, being interrupted only when a decision is owed, and being able to make that decision safely from a small screen.
What people do today
tmux + SSH. Keep the session alive on a box you can reach, SSH in from a mobile terminal app. Works, but you are reading a desktop TUI through a keyhole, and approving means typing into the same prompt the agent uses.
Notifications only. Hooks or wrappers that ping you when the agent needs input. You learn that it is waiting, but not what it wants to do, and you still need the laptop to review the diff before saying yes.
Cloud sandboxes with web UIs. Closest to the goal — but most render a chat transcript, not a decision. Scrolling three hundred lines of agent narration on a phone to find the one question is not reviewing; it is archaeology.
What a purpose-built agent terminal shows instead
VIVUU's answer is an attention queue rather than a transcript. Each waiting task arrives as a decision-ready card: what changed, why it matters, what the agent already did, and the one decision you owe — with the evidence a tap away.
Approval is press-and-hold — deliberate on purpose, sized for a thumb. Evidence (the diff, the test output, the staging log, the metric trend) opens full-screen with a checksum. Denying requires a written reason, and every decision lands in an append-only audit trail. The whole interface is portrait-first because the point is to work from the device already in your pocket.
Watch it run, not just wait
Between decisions, a task screen streams the agent's live events — tool calls, notes, evidence as it is produced — beside the plan with each step checked off. That is the difference between monitoring and refreshing: you can glance at a running task the way you glance at CI, then put the phone away knowing it will only interrupt you for a decision that is actually yours.
FAQ
- Can I monitor Claude Code remotely?
- Yes, with workarounds: keep the session alive in tmux and SSH in from a mobile terminal, use OS-level notification forwarding, or run the agent in a cloud sandbox with a web UI. What these lack is the decision layer — seeing the diff, the test run, and approving or denying with one deliberate action. That is the gap a dedicated agent terminal closes.
- What should a Claude Code manager UI actually show?
- Five things per task: what the agent is doing right now (live event stream), the plan and how far it has gotten, what it changed (diffs and test output as evidence), what decision it is waiting on, and how long that approval stays valid. Token or usage counters are secondary — the decision-blocking state is what strands agent work.
- Why do background coding agents stall when I step away?
- Most agent harnesses stop at permission prompts: a deploy, a destructive command, a file outside the sandbox. If the prompt only lives in a terminal window on your desk, the agent waits — sometimes for hours — until you return. Remote approval turns that dead time into a 20-second decision from wherever you are.
- Is a phone screen enough to review agent work responsibly?
- For the decision, yes — if the interface is built for it. You are not re-reading the whole codebase on a phone; you are checking the exact action, the target, the evidence (diff, test results, staging verification), and the blast radius. A portrait, decision-first layout makes that reviewable in under a minute; a shrunken desktop dashboard does not.
see the whole flow, live
Everything on this page — the queue, the evidence, the hold-to-approve, the audit trail — is working in the VIVUU terminal as an interactive preview on realistic mock data. No sign-up.