ukanwat/overtime
Give a coding agent a brief, not a chat, and it works on its own across sessions. Includes an example run: an open-world city built in a real game engine with no human help. In early development.
Overtime – Persistent Coding Agents
What it is
Overtime is a Node‑JS tool that lets you create agents – autonomous assistants that keep working on a task without you having to keep a session open. You brief an agent once in plain language, and it:
- plans the work,
- runs code‑generation models (Claude, Codex, Gemini, etc.) via the Agent Client Protocol (ACP),
- stores its own notes, memory and scripts in a dedicated folder,
- wakes itself on a schedule or when a watch‑script detects a trigger, and
- only pings you when a decision that requires your input is needed.
It is essentially an orchestration layer for existing LLM‑based coding agents, giving them a persistent, budget‑aware runtime.
Quick start
npm install -g @ukanwat/overtime # install the CLI globally
overtime # launch the interactive setup
- Give the new agent a name.
- Choose its runtime settings (backend model, daily budget, working folder).
- Describe the job in plain text – the agent writes an
AGENT.mdand starts.
Requires Node 22.19+ and a signed‑in coding agent (Claude works out‑of‑the‑box).
Core concepts
| Concept | What it means |
|---|---|
| Agent folder | ~/overtime/agents/<name>/ containing AGENT.md (task description), notes, memory files, and helper scripts. |
| Self‑time | Each agent decides when to wake: at a fixed time, on a cron‑like schedule, or via a watch script that triggers on file changes, log events, etc. |
| Budget & sandbox | A daily token/compute budget is enforced per agent; file‑system permissions keep the agent read‑only on protected paths. |
| Ask‑and‑continue | When a decision is required, the agent sends a message with options and its recommendation. You answer at your convenience; the agent resumes automatically. |
Example agents (pre‑canned use cases)
- repo‑keeper – monitors CI, auto‑fixes broken builds, asks before merging.
- on‑call – watches logs/metrics, diagnoses incidents, rolls back, writes post‑mortems.
- reviewer – reviews PRs, splits large reviews across helper agents, reports findings.
- release – creates releases, generates changelogs, runs canary deployments, asks before full rollout.
- research – periodically reads papers or docs, summarises only when something noteworthy appears.
- Trips, House hunt, Money, Learning, Side project – personal‑life agents that automate travel planning, rental hunting, budgeting, daily drills, and weekend‑project maintenance.
How it works under the hood
- Folder‑based agents – All state lives in plain files you can inspect or edit.
- ACP backend adapters – Overtime talks to any LLM‑coding service that implements the Agent Client Protocol (Claude, Codex, Gemini, OpenCode, etc.).
- Background daemon – After you close the CLI, a lightweight background process keeps the agents alive, waking them only when needed, so they incur no cost while idle.
- Watch scripts – Small user‑provided scripts can trigger an agent instantly (e.g., a new log line, a price drop, a new PR).
- Budget enforcement – The daemon tracks token usage per agent and stops further calls once the daily limit is hit.
Installation & requirements
- Node ≥ 22.19
- npm (global install of
@ukanwat/overtime) - At least one supported coding agent signed in (Claude works out‑of‑the‑box; Codex and Gemini need their ACP adapters).
Current status
- Early prototype – Claude backend is the most tested; Codex & Gemini adapters are present but less vetted.
- Planned features: mobile answering, quiet‑time tidying, morning stand‑up summaries, agent sharing, server‑side deployment for 24/7 watches.
License
MIT – free for personal and commercial use.
In a nutshell: Overtime gives you a persistent, budget‑controlled “assistant” that can run LLM‑based coding agents continuously, only interrupting you when a human decision is required. It’s aimed at developers who want their AI helpers to keep the wheels turning while they focus on higher‑level work.
Related
- Project
- Project
- Project
- Project
- Project