PromptPilot: a task scheduler for Claude Code and Codex that works while you sleep

Token quotas, request limits and reset windows are dictated by the provider, not the developer. But one resource stays under the developer's control: when a task is sent. That is what PromptPilot handles. It is a prompt-queue scheduler for AI CLI tools that lets you launch a code agent ahead of time, without you sitting at the monitor.

What PromptPilot is and why you might need it

PromptPilot is an open-source tool that one developer wrote for himself but published in the expectation that he is not the only one with this problem. The scenario is simple: you want to start a task on Claude Code or Codex at 1 a.m., when the provider resets your limit, but you are asleep by then. Or the reverse: in the morning you open your laptop and want the session warm-up to have already happened without you.

The system accepts tasks from three interfaces: a regular terminal, a web interface and a Telegram bot. A worker takes them from the queue and runs them one after another. If the tool hits a rate limit, it does not crash. It pauses and retries automatically. This matters most for people on free or limited plans, where every limit reset otherwise requires a manual restart.

Architecture: SQLite and three processes

The technical side is deliberately minimal. The entire queue is stored in SQLite, with no external brokers such as Redis or RabbitMQ. Three processes (an API server, a worker and a Telegram bot) communicate through the database. This lets you run everything on an ordinary laptop or a cheap VPS without complex infrastructure.

Tasks have priorities and start times. You can queue a task marked "run at 03:00" and the worker will pick it up at the right moment. Claude Code, Codex and Qwen are supported, which in practice means any CLI tool that accepts prompts from standard input. Higher-priority tasks run before the others, even if they were queued later.

The Telegram bot as a remote control

A separate use case is interactive communication with the model through the bot. If the AI agent asks a clarifying question mid-task, the user gets a Telegram notification and can reply right there, without opening a terminal. This is critical if you queued the task before leaving home, or if the agent stalled waiting for input while you slept.

The bot does not just show task status; it works as a full two-way channel to the model. For people who often work on the go or from a phone, that is a significant detail: you keep control of the agent even when you are physically far from your computer.

Why it matters

PromptPilot is not a product from a big company, and it is not a SaaS with a paid subscription. It is a tool built around one specific pain: the mismatch between a developer's working hours and the schedule of provider limits. The author admits he built the project for himself, but that is exactly why it solves a real problem rather than an invented one.

As AI CLI tools become a standard part of the workflow, the need to manage task flow and limits will only grow. PromptPilot fills that gap with simple means: no Kubernetes, no message queues, no complex deployment. One SQLite database, three processes and a Telegram bot, and your AI agent runs on a schedule.

The project's code is published, and deployment is described in the README. This is a case where open source solves the very problem commercial tools overlook, because it is too small to justify building a separate product around.