Phone approvals

Your agent asks.
Your phone answers.

When an AI agent wants to run something risky — rm -rf, a force-push, a production migration — the question lands on your lock screen. One tap approves or denies it, and the agent resumes in about three seconds. You can be at lunch.

claude — ~/api
$ claude "clean up the build system"
● Reading Makefile, scripts/…
▶ Bash rm -rf ./build
Paused — approval requested on your phone
✔ Approved from phone · resumed in 3s
✖ Denied from phone · this call blocked, agent continues
12:47
Saturday, July 26
ClawMetrynow
Approval needed: Bash
Policy 'rm_rf' fired: rm -rf ./build
✅ Approved — your agent has been told.
❌ Denied — just this call, session lives.
↑ try it — tap a button

One tap. About three seconds.

A real round-trip, not a page you have to go find: the decision travels from your lock screen back to the paused agent.

1

A policy fires

Your rules watch every tool call — destructive deletes, force pushes, sudo, DB mutations, secrets, package installs, network calls. Presets exist for all of them; enable the ones you want.

2

Your phone buzzes

The notification carries the question itself: which tool, which policy, the command preview. Android answers straight from the notification; iPhone opens the approval in one tap.

3

The agent resumes

Your decision is picked up within ~3 seconds. Approve and it proceeds — no second prompt at the desk. Deny and that call is blocked. Unanswered requests time out to the action you chose (deny by default).

Built like a gate, not an alarm

Most notifiers tell you what already happened. This one stops the action first and asks.

PreToolUse

Blocks before execution

For Claude Code, clawmetry hooks install wires a pre-execution hook: the tool call pauses before it runs and waits for your decision. Not detect-and-regret — the rm -rf never happens unless you say so. OpenClaw gets the same via its exec-policy gate.

deny ≠ kill

Deny one call, keep the agent

Denying from your phone blocks that single tool call. The agent keeps its session, its context, and its progress — it's told why and tries another approach. No more killing an hour of work to stop one command.

E2E posture

The question travels, not your code

The push carries the tool name, the policy, and a short command preview — not your source, not your transcripts. Web Push encrypts the payload in transit, and decisions ride signed one-time links.

Two minutes, twice

Once on your phone, once on the machine your agents run on.

📱 Your phone

No app store. Works on Android (Chrome) and iPhone (iOS 16.4+).
  1. Open app.clawmetry.com/push on your phone
  2. iPhone only: Share → Add to Home Screen, open from the icon
  3. Tap Turn on notifications, then Send a test
Lock your phone before the test for the full effect.

🖥 Your agent machine

Connects the daemon and installs the Claude Code pre-execution gate.
$ pip install -U clawmetry $ clawmetry connect $ clawmetry hooks install
Then flip on the policy presets you want in the dashboard's Approvals tab — they're off until you choose them. clawmetry hooks uninstall reverts cleanly.

Approval policies watch every runtime ClawMetry supports:
OpenClaw · NVIDIA NemoClaw · Claude Code · Codex · Cursor · Aider · Goose · opencode · Qwen Code · Hermes · PicoClaw · NanoClaw · Pi · Deep Agents

Questions

Which phones does it work on?

Android (Chrome): Approve/Deny buttons appear directly in the notification — answer from the lock screen without opening anything. iPhone (iOS 16.4+): add ClawMetry to your Home Screen once (Safari → Share → Add to Home Screen); a tap on the notification opens the approval with big Approve/Deny buttons. No app-store install on either.

What exactly happens on deny?

With the Claude Code hook installed, deny blocks that one tool call before it executes — the session survives and the agent is told to try something else. Without the hook, a denied approval stops the offending session instead.

What if I don't answer in time?

Every policy has a timeout (default 60s) and an on-timeout action you choose. The default is deny — an unanswered risky call never slips through by silence.

What reaches ClawMetry's servers and my phone?

The question: tool name, policy name, and a short command preview. Your code and transcripts aren't part of it. Web Push encrypts payloads in transit; decisions use signed single-purpose links.

Does this need the paid plan?

Phone push and the approvals queue are part of ClawMetry Cloud. The free plan covers OpenClaw and NVIDIA NemoClaw agents; watching Claude Code, Codex, Cursor and other paid runtimes needs the Starter plan — free trial, no card, cancel anytime. See pricing.

Can I also get "agent finished" pushes?

Yes — the same hook install adds a notification when Claude Code finishes and is waiting on you, or when it raises a permission prompt your policies didn't cover.

Stop babysitting the terminal.

Enable it now — the next time an agent needs you, you'll know wherever you are.

Enable on your phone →