claudex

claudex

Two AIs argue about your plan so you don't have to.
Claude turns your one-line idea into a detailed plan. Codex (a different AI) grills it from three different angles. Claude revises. They keep going until there's nothing left to fix. You watch from one terminal window.

_{The GitHub repo is the v1 teachable artifact. The actively-maintained build with new features, fixes, and supporting workflows lives in the Early AI-dopters community.}

---

What you actually get

• A plan that's already been argued with. You start with a one-liner. You walk away with a numbered plan that has survived rounds of cross-examination.
• Two AIs cross-checking each other. Claude is the writer. Codex is the reviewer. Neither one gets the last word alone.
• Three different reviewer hats. Round 1 thinks like a senior engineer. Round 2 thinks like a security person. Round 3 thinks like an ops person. So you don't miss a category of problem.
• One window, no babysitting. You type one slash command. Walk away. Come back to a finished plan and a clean summary of what changed each round.

Who this is for

Claudex is a developer tool. You'll be comfortable with it if:

• You already use Claude Code (Anthropic's official CLI).
• You're okay running a few terminal commands to install dependencies.
• You have a ChatGPT Plus, Pro, Team, or Enterprise account (Codex authenticates against that).

If "terminal", "slash command", or "CLI" feel foreign, you'll get more value from the Early AI-dopters community, where this stuff gets taught with hand-holding instead of from a README.

---

What it looks like in practice

$ /claudex:plan add expiry dates to my links

  Want me to interview you to sharpen the topic, or just go?
    → Interview me first
    → Just launch the loop

  Round 1 of 3, Senior-engineer review
        Codex: 5 findings (2 high, 3 medium)
  Round 2 of 3, Security and data-integrity review
        Codex: 1 high, 1 low
  Round 3 of 3, Ops and SRE review
        Codex: no substantive findings
  LGTM. Plan locked.

  Plan: PLAN.md
  Log:  .claude/claudex/<id>.log

You typed one command. You watched three different reviewer personas grill the plan from three different angles until there was nothing left to grill. You walked away with a vetted plan. You did not touch the keyboard between the first command and the final output.

That's claudex.

---

Under the hood

Two slash commands wired through a Claude Code Stop hook, plus four utility commands. The Stop hook is the only mechanism in Claude Code that can force an autonomous loop. Claudex uses it to drive Claude and Codex back and forth until the work is done.

Command	Mode	Behavior
/claudex:plan [flags] <feature>	Plan mode	Optionally interviews you to sharpen the topic, then Claude drafts PLAN.md, Codex pressure-tests it, Claude revises. Each round uses a different reviewer persona. Loops until LGTM or N rounds.
/claudex:review	Review mode	Codex reviews the diff. Findings + proposed fixes written to reviews/. Read-only in v1.
/claudex:status	utility	Print the current loop's mode, phase, round, elapsed time, and per-round severity tallies. Read-only.
/claudex:doctor	utility	Preflight diagnostic. Verifies bash, codex CLI, plugin file integrity, hook fail-open. Run after install.
/claudex:cancel	utility	Graceful cancel of the active loop.
/claudex:rollback	utility	Nuclear cleanup of all state files.

Plan-mode flags