This repository now ships a parallel Codex marketplace alongside the existing Claude Code marketplace.
Codex loads repo marketplaces from .agents/plugins/marketplace.json and expects each installable plugin to provide a manifest at .codex-plugin/plugin.json.
Current Codex plugin docs describe these plugin surfaces:
skills/.mcp.json.app.json
That means this repository can expose skill-bearing plugins to Codex today, but the hook-only guard plugins remain Claude-specific until Codex exposes equivalent plugin components or those behaviors are repackaged behind skills or MCP servers.
The repo marketplace at .agents/plugins/marketplace.json currently includes:
corepython-backenddeep-research-pluslinux-hyprland
These plugins now ship .codex-plugin/plugin.json manifests, and the plugins that need MCP setup also ship .mcp.json.
For subagents specifically, the repo also generates project-scoped Codex custom
agents under .codex/agents/ from the Claude agent markdown. Those files are
generated artifacts, not hand-maintained copies.
For Claude slash commands specifically, the repo also generates project-scoped
Codex custom prompts under .codex/prompts/ from the Claude command markdown.
Those prompts are invocable as /prompts:<name> in Codex, for example
/prompts:workflows-review.
Regenerate both with:
uv run python scripts/generate_codex_agents.pyInstall them into your Codex home with:
./scripts/install-codex-agents.shThis installs generated agents into ~/.codex/agents/ and generated prompts
into ~/.codex/prompts/.
Do not hand-maintain a second Codex copy of the instruction tree.
The repo now supports a generated Codex workflow from the canonical Claude plugin source:
uv run python scripts/generate_codex_agents.pyThat Codex workflow:
- generates
.codex/agents/*.tomlfrom Claudeagents/ - generates
.codex/prompts/*.mdfrom Claudecommands/ - enforces collision checks and blocks reserved built-in Codex agent names
- rewrites common Claude
Task(...)/AskUserQuestionpatterns into Codex agent phrasing - rewrites known Claude slash-command references like
/workflows:reviewinto Codex prompt references like/prompts:workflows-review - carries plugin-level MCP configuration into generated agent files when present
Install the generated Codex artifacts with:
./scripts/install-codex-agents.shThis keeps the source of truth in the Claude markdown and emits Codex artifacts mechanically.
OpenAI's Codex docs currently describe custom prompts as deprecated and recommend skills for reusable prompts instead. This repository still generates custom prompts because they remain the closest Codex-native equivalent to Claude's explicit slash-command entrypoints.
For long-term shared workflow packaging, Codex skills remain the more stable surface. For command-like explicit invocation parity, custom prompts are still useful today.
These plugins are not listed in the Codex marketplace because they are primarily hook plugins:
enforce-uvconventional-commitsgemini-model-guardpython-formatpython-typechecktest-reminderclean-code-guardsafety-guardgit-safety-guardgh-api-guardprotect-env
They still work through the Claude marketplace and .claude-plugin manifests.
When adding a new plugin that should work in both ecosystems:
- Keep the Claude manifest at
.claude-plugin/plugin.json. - Add a Codex manifest at
.codex-plugin/plugin.json. - Add the plugin to
.agents/plugins/marketplace.jsononly if it has meaningful Codex-native surfaces such asskills,apps, ormcpServers. - Keep Codex paths relative to the plugin root and start them with
./. - If the plugin is hook-only, leave it Claude-only until there is a real Codex equivalent.