NightForge — Let Claude Code and Cursor Work Autonomously at Night

Context

NightForge was born out of a developer’s frustration. With a Claude Max subscription, a coding agent like Claude Code operates within quota windows: once the 5-hour window is exhausted, everything stops, and you must manually type “continue” to restart once the quota recharges. At night, this time is simply wasted.

NightForge runs Claude Code and Cursor autonomously on a developer’s computers, processing pre-prepared prompt queues, and is controlled via a phone.

Problem

• Quotas recharge at times when no one is at their screen
• Chaining multiple tasks across projects requires constant supervision
• Running an agent remotely requires exposing a PC (VPN, open port)
• Upon waking, you still need to track modifications and their locations.

Solution

Developers structure their night project by project: a sequence of messages, the tool (Claude Code or Cursor), the model, and the target machine. A scheduler calculates when each quota will be available and automatically resumes work after each recharge.

On each PC, a Python agent establishes an outbound connection to the central server—no open ports, no VPN. It launches tools via command line using the developer’s subscriptions, pushes code to GitHub (to the main branch or a date-stamped nightly branch), and streams logs in real-time.

The same Nuxt interface is served as a web app, a mobile PWA, and a Tauri desktop app for Windows, with automatic updates.

Key features

• Project-based prompt queues (tool, model, effort, target machine)
• Night composition with a reusable prompt library
• Claude Max quota scheduling and automatic resumption after recharge
• Simultaneous execution of Claude Code and Cursor
• Parallel execution across multiple machines
• Code review: modified files and line-by-line diffs
• Phone-based discussions with push notifications
• Safeguards: emergency stop and auto-stop after repeated failures.

Results

NightForge is live at nightforge.dibodev.fr, with a continuously updated desktop app (over 30 versions in a few weeks). It helps advance projects overnight, and the code is publicly available on GitHub.

The project embodies a simple idea: an AI agent is only valuable if properly orchestrated, with clear tasks, safeguards, and human oversight upon waking.