Can Claude Code run overnight?

Yes — Claude Code can run multi-hour autonomous sessions, but only if the host machine stays alive. Laptops that sleep will halt the run mid-execution. Use an always-on host (Mac mini, Windows VPS, or managed cloud desktop).

How long can a Claude Code session actually run?

Practical session length depends on the complexity of the task and Anthropic's per-request token limits. Long refactors typically run 1–6 hours of wall-clock time when broken into multi-phase plans. Standby customers regularly complete 4-hour refactor sessions overnight.

What if Claude Code gets stuck overnight?

Build a 'safety prompt' into the kickoff: instruct Claude Code to commit work-in-progress every 30 minutes and write a status note. If it's stuck the next morning, you have a partial branch + the status log to diagnose. Also configure Claude Code to ask before making destructive changes.

How to run Claude Code overnight — the complete playbook

The short version

Three things have to be true to run Claude Code overnight successfully:

Your host stays alive. Laptops that sleep will halt the run mid-execution. Use a Mac mini, Windows VPS, or managed cloud desktop.
Your prompt is structured for long autonomy. Break the task into phases. Tell Claude Code to commit every 30 minutes. Ask before destructive changes.
Your safety net is real. Run on a feature branch, never main. Have tests that auto-run. Set spending limits.

Get those three right and a 4-hour overnight refactor is one of the highest-leverage things you can do as a developer in 2026.

Step 1 — Pick the host

Anything that runs Claude Desktop. Options ranked by reliability for overnight runs:

Managed cloud desktop (Standby, $29/mo) — 99.9% uptime, you don't think about it. Best for overnight runs.
DIY Windows VPS (Kamatera, ~$15/mo) — 99.5% uptime, you handle OS patching.
Mac mini at home — 95% uptime, depends on your home Wi-Fi and power. Fine if Wi-Fi outages and reboots are rare.
Laptop kept on — 85% uptime. Workable but expect 1–2 missed runs per month.

Whichever you pick, configure: power policy to never sleep, Claude Desktop to auto-launch at login, auto-login enabled, Windows Update active hours wide.

Step 2 — Structure the prompt for autonomy

The single biggest difference between an overnight run that finishes vs. one that hangs is how the kickoff prompt is structured. The bones of a good one:

You are going to refactor <target dir> while I sleep.

CONTEXT: <1-paragraph summary of what the code is and why>

GOAL: <1-paragraph specific outcome, e.g. "extract all shared
hooks into src/hooks/, dedupe duplicate utils, add types where
missing, leave tests green">

PLAN: Generate a multi-phase plan before starting work. Show me
the plan in your first response. Do not start coding until I
confirm — oh wait, I'm asleep. Just proceed if the plan looks
reasonable to you, but commit the plan to a PLAN.md at the
root of a new branch before you start.

WORKFLOW:
- Create a new branch named refactor/auto-<today>
- Commit every 30 minutes with a clear message
- Run tests after each phase; if tests fail, fix before
  moving on
- Write status updates to STATUS.md every hour
- Never push to main, never delete files outside <target dir>,
  never modify package.json without noting it in STATUS.md

WHEN DONE: open a PR against main with a summary of changes,
tag <your handle> for review

START: now

This structure is doing several things: (1) gives Claude permission to proceed without you, (2) forces a plan-first workflow so you can debug if it goes wrong, (3) sets safety rails (branch, commit cadence, no destructive ops), (4) leaves a paper trail (STATUS.md, PLAN.md) for the morning.

Step 3 — Safety nets

Three you should never skip:

Branch-only commits. Never let Claude Code push to main. Always feature branch with a PR.
Tests on every commit. If your repo has CI, fine. If not, add a pre-commit hook that runs unit tests.
API spending limit. In Anthropic's console, set a monthly spend cap. A runaway loop can burn budget overnight.

Step 4 — The morning routine

When you wake up:

Open Claude Desktop on your host (via RDP, Dispatch, or just walk over to the Mac mini). Confirm the agent is alive — not crashed, not at a "are you sure?" prompt.
Check STATUS.md on the branch. If Claude Code is hung, this tells you where.
Open the PR. Review the diff at high level. Run tests locally if there's no CI.
If it looks good: approve, squash, merge.
If it doesn't: ask Claude Code via the same session to fix, or pull the branch locally and finish by hand.

Things that go wrong (and how to fix)

Host slept overnight. Tasks halt. The whole thing fails silently. Fix: use a host that doesn't sleep. (This is why we built Standby.)
Claude Code asks a clarifying question and waits. 8-hour timeout, no response. Fix: kickoff prompt should grant autonomy ("just proceed if the plan looks reasonable to you").
Tests fail and Claude Code stops. The "fix tests before moving on" rule sometimes loops. Fix: cap with "if tests still fail after 3 attempts, commit progress as WIP and move to the next phase, note it in STATUS.md."
API rate limit / spending cap hit. Run stops cold. Fix: raise the cap or schedule the run for off-peak hours.
Windows Update reboots at 3 AM. Fix: set Active Hours to all day, or use a managed service that handles this.

An honest expectation

A well-prompted overnight Claude Code run finishes successfully about 70% of the time today (May 2026). 20% finishes with caveats (some tests failing, partial scope, need to clean up by hand). 10% fails entirely (gets stuck, runs out of budget, host crashes). Those numbers are improving fast as Anthropic ships better long-context agentic behavior.

Even at 70% success, "I shipped a 47-file refactor while sleeping" is one of the highest-leverage moments developers have access to in 2026. The infrastructure side is solved. The remaining variance is in your prompt.

Want this hands-off?

If you're tired of managing the host, Standby is the managed version — a Windows cloud desktop with Claude Desktop, Claude Code, Chrome, Git, and Node pre-installed. You sign in with your own Anthropic account in 10 minutes. We patch the OS, we keep Claude Desktop launched, we monitor uptime. The overnight runs work because the desktop never sleeps.