Skip to content

cross-review installs skills for Claude Code and Codex, and registers its MCP server with both. The Codex CLI is the reviewer, so it must be installed too. One command does the whole install.

  • Node.js 22 or later and npm.
  • Claude Code CLI, Codex CLI, or both. The skills work in whichever client you use.
  • codex in your PATH, logged in with an OpenAI account or an API key. Codex is the default reviewer. Install it with npm install -g @openai/codex if it is missing.
Terminal window
npx @tlelabs/cross-review

To install the default branch of the GitHub repository instead of the npm release:

Terminal window
npx github:lploc94/cross-review
  1. Checks that Node.js is version 22 or later.
  2. Copies the shared runtime (the runner, the MCP server and the dashboard) into ~/.claude/skills/cross-review/ and installs its dependencies from the lock file.
  3. Fills in each skill’s template with the shared review protocol and the absolute path of the runner.
  4. Runs the runner’s version command to confirm it works.
  5. Replaces the skill folders in ~/.claude/skills/ and ~/.agents/skills/ in one step, and puts the previous files back if anything fails. It also removes the old codex-review pack; see Move from codex-review.
  6. Registers the cross-review MCP server with Claude Code (user scope) and with Codex. The Codex entry gets a 300-second tool timeout.
  7. Writes the cross-review command to ~/.local/bin/cross-review on macOS and Linux.

The installer prints the skills, the MCP registrations and the command it wrote. Re-running it is safe: the MCP registration is replaced, not duplicated.

Check the runner version. It prints 15:

Terminal window
cross-review version

If the shell cannot find cross-review, the folder ~/.local/bin is probably not on your PATH. You can always run the runner by its full path:

Terminal window
node ~/.claude/skills/cross-review/scripts/review-runner.js version

Check the MCP registrations:

Terminal window
claude mcp get cross-review
codex mcp get cross-review

When a registration is missing, see MCP server is not registered. Then run a review as shown in the Quickstart.

Terminal window
npx @tlelabs/cross-review@latest

The installer replaces the runtime and the skills. Let a review that is still running finish first, since it runs on that runtime.

Terminal window
npx @tlelabs/cross-review --auto

--auto adds a block to ~/.claude/CLAUDE.md. The block tells Claude Code to run review-plan after it writes a plan and review-impl after it finishes an implementation, without asking first. Every install refreshes a block that is already there. The block is described in Automatic reviews.

  • The installer does not write ~/.local/bin/cross-review. Start the dashboard with npx @tlelabs/cross-review dashboard instead of cross-review dashboard.
  • Skills, MCP registrations and the runner are installed the same way as on macOS and Linux.
  • For the other runner commands, run node on ~/.claude/skills/cross-review/scripts/review-runner.js with the command as its argument.

The earlier pack was called codex-review, with skills named codex-plan-review, codex-impl-review, codex-think-about, codex-commit-review and codex-pr-review. The installer removes:

  • the old codex-* skill folders and the codex-review runtime, when they belong to that pack;
  • the codex-review MCP registration, when it points at that pack’s runtime;
  • copies of this pack’s skills in $CODEX_HOME/skills, Codex’s older skills folder.

Skills the installer did not put there are left alone. The review-commit skill was removed in version 15; review-impl now covers its job, reviewing unpushed or recent commits.