Autoprompt 🤖 Agent Open source

What it does

Autoprompt, by Spielewoy, installs a structured multi-agent workflow into an existing coding agent. You give it a goal with constraints and a way to check success, and it runs the loop for you: choose a route, plan, build, check with independent reviewers, repair, and finish. The idea is that the agent doing the work should not also be the one approving it, so planning, execution and verification are split across separate roles.

What is inside

The package is a small CLI called autoprompt that writes agent definitions and instructions into each supported tool. From there you start runs with autoprompt activate PROVIDER -- "<goal>". Runs can take one of four paths: direct for focused work, light with a short plan, roadmap for dependent multi-step work, or auto to let it pick. Concurrency controls cap how many subagents run at once, and a configure command lets you pin a model or effort level for the worker agents. A doctor command checks installations and an uninstall command removes them. The README reports Terminal-Bench gains from its first version; those are the author's own measurements.

Works with

The support table lists Claude Code, Codex, OpenCode, Kilo Code and VS Code agent mode as tested, alongside several other harnesses. A custom-agent guide covers tools that are not on the list.

How to install

npm install -g https://github.com/Spielewoy/autoprompt-skill/releases/download/v2.0.0/autoprompt-skill-2.0.0.tgz
autoprompt

Maintenance and safety

MIT licensed, recently released and pushed, with contributing, security and code-of-conduct files. It needs Node.js 20+, Python 3.11+ with PyYAML, and Bash 4.3+, so Windows users need WSL. Runs are started explicitly, never in the background, and it stops when a choice needs your authority. Multi-agent runs use more tokens than a single session.

Who should use it

Developers who hand whole features or bug fixes to a coding agent and want a built-in review and repair loop instead of prompting every step. Try it on a contained task first to judge the extra cost.