wtf is coder?

Coder is a tool that lets you hand off coding work to separate AI subagents, regardless of which AI assistant you normally use or which AI model powers those subagents. Instead of one AI doing everything in a single long conversation, Coder splits off implementation details into supervised subagents while your main conversation stays focused on intent and coordination. It works across different hosts, meaning it plugs into tools like Claude Code, Codex, and anything that reads the shared Agent Skills folder on your computer. It is also cross model, so a task can be handled by Codex, Claude, another OpenAI compatible service, or even a model you run locally, and you can choose which model handles which task. Handoffs to subagents are meant to be quick and lightweight, so you can spin up many of them at once without slowing things down. Every task dispatched to a subagent stays visible: you can watch its progress live, correct its course partway through, or stop it entirely. Permissions are unified across every underlying engine into three simple modes: read only, workspace write, or fully automatic. Getting started means installing Coder into whichever host you use, such as Claude Code, either through that tool's own plugin system or through a global npm install. You then need at least one connected engine, which usually means logging into an existing AI coding subscription like Codex or Claude Code, though you can also point it at your own model. Once set up, you simply ask your regular AI assistant to use Coder for a task, or set it as the default for implementation work in your project's instructions file. Configuration lives in a JSON file, either machine wide or per project, where you define which engines to chain together and which model and permission level each one uses. A command line interface is also available for anyone who wants to dispatch, monitor, steer, or stop tasks directly rather than through their AI assistant.