Codex

Where rmk puts each type of item for Codex, and what Codex asks of you first.

From the Documentation in Ronne AI Marketplace 0.3.0.

Where each type goes

With --target codex, rmk writes each item where Codex reads it: in the project, or in your home folder with --scope user.

Type	Where it goes
`skill`	.agents/skills/<name>/ (your home folder: ~/.agents/skills/)
`agent`	.codex/agents/<name>.toml
`rule`	a section in AGENTS.md (your home folder: ~/.codex/AGENTS.md), or a skill when the AI decides or you ask
`command`	.agents/skills/<name>/, run as /<name>
`hook`	hooks in .codex/hooks.json; a script under .codex/hooks/<name>/
`mcp-server`	mcp_servers in .codex/config.toml
`permission-policy`	.codex/rules/<name>.rules, for shell commands only
`output-style`	not supported: skipped with a warning
`statusline`	not supported: skipped with a warning
`lsp-server`	not supported: skipped with a warning
`bundle`	nothing of its own: its items are installed one by one

rmk export --from codex reads the agents and MCP servers you wrote for Codex from the same places: Exporting your own items.

Trust and hook review

Codex reads a project's .codex/config.toml, hooks and rules only once you trust the project, and runs new or changed hooks only after you review them with /hooks. rmk reminds you after an install that needs either.

Good to know

Skills and commands go to .agents/skills/, the folder Cursor reads too, so one copy serves both. Commands become skills you run as /name; Codex doesn't pass them arguments, so a note in the skill says what each placeholder is.
Rules are sections of AGENTS.md between rmk:begin and rmk:end markers, and your own text around them is never touched. Codex reads at most 32 KiB of instructions, and rmk warns when AGENTS.md passes that.
MCP servers get their secrets by the variable's name (env_vars, bearer_token_env_var): nothing secret is written, and you set the variables yourself.
rmk writes config.toml back in one layout when it adds or removes a server, so comments in it aren't kept. It says so the first time.
Codex agents have no tool list and no fast or strong model: an agent gets the session's tools and model, unless its item sets a Codex model. Hooks run for every tool, since Codex's tool names aren't documented; a hook meant for one tool reads the event on stdin and checks.
Permission policies become Codex rules for shell commands only, which Codex marks as experimental. Output styles, status lines and language servers have no place in Codex, and are skipped with a warning.

Plugins

Codex can also install this marketplace's items as plugins, from a git repository that rmk feed build keeps as a mirror of it. Codex plugins carry skills, MCP servers and hooks, not agents. Plugin marketplaces explains how.