The short answer
Two parts: let it run, then choose how the news reaches you.
Hands-free agent work has two parts: letting the agent run without approving every step, and finding out what happened without watching it. For the first, give the agent a bounded task with explicit permissions and stop conditions, run Claude Code non-interactively with claude -p, or hand the task to a cloud agent. For the second, match the channel to how far away you are. A cloud agent’s pull request or an emailed result suits checking back later. A phone push from Claude Code Remote Control, or a desktop alert, tells you when a run finishes or needs a decision. Spoken updates from Heard on your Mac suit being nearby with your eyes elsewhere, when you want to hear progress, questions and failures as they happen.
Whatever the channel, an update is not authorization. Deployment, deletion, credential changes and unreviewed code still get a direct look before you accept them.
Make success and stopping conditions explicit.
A hands-free task needs a contract: the deliverable, the files or subsystem in scope, the tests to run, the actions that require approval, and what evidence the agent must report at the end. Good candidates are bounded and reversible: a test run, a contained refactor, documentation, a dependency analysis.
Headless mode removes the prompts, not the risk. claude -p runs Claude Code without the interactive UI, and --allowedTools lists the tools it may use without asking. It also tells you nothing unless you wire up an output, which is what the next sections are for.
- No destructive actions without approval.
- No expansion into adjacent systems.
- Stop after repeated failure or unclear requirements.
- Report modified files and verification results.
Goal: Refactor the date helpers in src/utils/date.ts. No behavior change.Scope: src/utils/date.ts and its tests only.Checks: run npm test -- date before you finish.Ask first: deleting files, changing dependencies, anything outside scope.Stop if: tests fail twice in a row or the requirement is unclear.Report: files changed, test output, anything left unverified.Updates after the fact: pull requests and emailed results.
If you will be gone for hours, let the result come to you as an artifact. In VS Code, cloud agents return pull requests, so the update and the review happen in the same place. With Claude Code, a script or queue can run claude -p, push the commit and email you the outcome; Sam French describes one such setup.
Where this wins: no machine has to stay awake for a cloud agent, and the pull request is already the thing you need to review. Where it is weak: you only find out when you check, so a question in minute five waits until you come back.
Updates the moment something changes: phone push and desktop alerts.
Claude Code Remote Control can push to the Claude mobile app. Start the session with claude --remote-control (or run /remote-control in an open session), then run /config and turn on Push when Claude decides (usually when a long task finishes or it needs a decision) or Push when actions required (permission prompts and questions). The session still runs on your computer, so it has to stay on, and Remote Control needs a Pro, Max, Team or Enterprise plan (API keys are not supported). For Codex in the ChatGPT desktop app, Settings lets you show turn-completion alerts never, only while ChatGPT is in the background, or always, and permission and question notifications have their own switches.
For hook-based setups on a Mac, see how to get notified when Claude Code needs your input and how to play a sound when it finishes. A push or banner is a one-line heads-up; for the detail you open the session.
Updates while you are nearby: spoken progress.
When you are in the same room but looking at something else, a voice carries more than a banner. Heard (heard.dev), the voice layer for coding agents, narrates Claude Code and Codex on your Mac in short spoken updates: what changed, what needs you, what failed, what finished. Pick how much you hear with its listening modes: Co-pilot at the screen, Companion when you step away, Focus when only a question, an approval or a failure should interrupt. On Heard Power you can answer by voice: hold right Command and start with the agent’s name.
Cloud agents can report to the same voice through Heard’s MCP connector, including the GitHub Copilot coding agent, Devin and Grok Bot, and so can Claude Code or Codex running on another machine. They are heard only when they send a report, because Heard cannot watch a cloud agent from outside.
- At the screen, reading alongCo-pilot
- Away from the screen, following a long task by earCompanion
- Heads-down on other work, speak only for a question, approval or failureFocus
Which channel for which distance.
Two channels usually cover it: one for when you are close and one for when you are not.
- Away for hours: a pull request or an emailed result.
- In another room or out: a phone push from Remote Control.
- At the desk, other window in focus: a desktop alert or a sound.
- Nearby, eyes elsewhere: spoken updates from Heard.
Review the artifacts, not just the update.
Read the diff, inspect any security-sensitive behavior, run or review test output, and confirm the change stays within scope. A confident summary, spoken or written, is not evidence by itself.
Safe workflow
Scope, listen, intervene, verify
Scope Assign a bounded, reversible task with checks and stop conditions.
Listen Choose the update channel that fits your distance, and the least verbose setting that still tells you about state changes.
Intervene Return for questions, failures, sensitive actions, or uncertainty.
Verify Inspect the diff and evidence before accepting the result.
Questions
Frequently asked questions
Can Claude Code send updates to my phone?
Yes, through Remote Control. Install the Claude app, sign in with the same account, start the session with claude --remote-control, then in /config turn on Push when Claude decides, Push when actions required, or both. The session runs on your computer, so it must stay on.
What is the difference between headless mode and hands-free updates?
Headless mode (claude -p) runs the agent without the interactive UI. It does not tell you anything by itself; you still need an output such as a commit, an email, a push or a voice.
How do I hear a cloud agent’s updates?
Connect it to Heard’s MCP server from Settings > Connections > Cloud agents. Heard speaks each report the agent sends; it cannot watch a cloud agent that does not report.
What should always bring me back to the screen?
Destructive actions, production changes, credentials, security-sensitive behavior, repeated failures, unclear scope, and final acceptance all deserve direct review.
Ready when your agents are
Hear the work without watching every screen.
Start with the full Heard Power experience for 14 days. No card required. The open source engine remains free.
Download Heard for Mac macOS 14 or later · Claude Code and Codex supported