Claude in Codex

Claude in Codex plugin logo

Run Claude Code from Codex for adversarial code review and task delegation.

Skills

claude-in-codex:adversarial-review: Invokes Claude Code as a review-scoped adversarial reviewer, has Codex triage the feedback, applies accepted in-scope fixes for current working-tree changes, validates the result, and summarizes the outcome.
claude-in-codex:using-claude-cli: Internal contract for non-interactive Claude Code CLI use — command shape, permission posture, sessions, structured output, and prompting guidance. The review skill and the Claude proxy subagent use it; it is not meant for direct invocation.

Claude as a Codex subagent

Codex plugins cannot package subagent definitions, so this plugin ships a copyable one instead. Copy skills/using-claude-cli/references/claude-agent.toml to .codex/agents/claude.toml in a project, or to ~/.codex/agents/claude.toml for all projects. The copied definition lets Codex spawn Claude Code as a general-purpose proxy subagent for delegated tasks.

The definition is a work in progress. Custom-agent support has two upstream blockers:

Symlinked custom-agent TOMLs are not discovered.
Tool-backed sessions cannot reliably select custom agents by name, including Codex app sessions whose subagent tool exposes no agent-role selector.

Direct copies can work in compatible Codex CLI sessions, but behavior varies by the subagent tool surface exposed to the session. Prefer the adversarial-review skill for its focused code-review workflow, or invoke the Claude CLI directly through the using-claude-cli contract.

The Claude CLI cannot run inside Codex's sandbox: it needs network access to reach the Anthropic API, and it needs home-directory writes for session state. The proxy therefore stays read-only, and every claude command requests escalated execution. Whether Codex grants that exception depends on the parent session's approval policy, reviewer, and rules. Live permission overrides on the parent session take precedence over the proxy's configured read-only default.

Requirements

claude must be available on PATH.
Claude Code must already be installed, authenticated, and ready to run non-interactively.
The Claude CLI needs network access and home-directory writes wherever it runs; from inside a Codex session, claude commands must run with escalated permissions outside the Codex sandbox.

Codex support

Codex loads this plugin from the extensions.com.openai block of the root plugin.json, the Agent Plugins 1.0.0 manifest.
Skill content lives in skills/ at the plugin root.
Codex-specific skill metadata lives beside each skill in agents/openai.yaml.