AGENTS.md

Write your project's instructions once, in AGENTS.md at the project root. Every agent reads it: Codex, OpenCode and Antigravity natively, and Claude Code through a one-line CLAUDE.md that uze keeps in place.

Set it up

Ask any agent, with the uze:init skill:

/uze:init

It reads your project and whatever instructions already exist (CLAUDE.md, GEMINI.md, …), proposes one AGENTS.md, and changes nothing until you agree.

Check it

uze status

Says which file your instructions come from and whether every agent reads it. If one does not, uze install fixes it.

What uze writes into it

AGENTS.md is yours. The only place uze writes is inside the blocks it marks:

<!-- uze:begin … -->
…
<!-- uze:end … -->

They hold what uze needs every agent to know. uze install writes the ones from your plugins and one on authoring plugins, for any project with an agents.yaml. The workspace writes its own, on how it isolates the agents it launches, only while you use it and only once the project chooses a policy (Agents); uze install never touches that one. Anything you write outside the blocks is never touched, and a block you edit by hand is reported, not overwritten.

Project skills and agents

Skills and agents the project keeps in .agents/skills and .agents/agents reach Codex, OpenCode and Antigravity natively, wherever you start them. Claude Code does not read .agents/, and Codex does not read project agents there, so for those two they arrive through the workspace's shim: inside uze workspace only. uze status and uze inspect say which. Plugins you install are not affected: they reach every agent everywhere.