How to use Claude Code and Codex together
The short answer
Use Claude Code to make a change, then ask Codex to review the same files through OpenScout. Save the returned reference to read the findings and follow up in the same review conversation.
bun add -g @openscout/scout- Claude CodeMakes the change and leaves it uncommitted in your checkout.
You run
scout ask --project … --harness codex - ScoutRoutes the request and hands you a reference. Example:
ref:1a2b3c4dStarts a fresh Codex session in the same checkout
- CodexReviews the same files and reports findings.
scout wait ref:… reads the findingsscout ask --ref ref:… asks a follow-upThis walkthrough uses terminal commands; no plugin is required.
1. Set up the local connection
Before you start
You need Claude Code and Codex installed and authenticated on the same machine, plus a Git project both can read. For this first run, use a trusted local development machine with Bun 1.3 or newer. The native Scout app requires macOS 26 or newer; this walkthrough uses the CLI.
Scout isn’t installed yet
If Scout is not installed, run:
bun add -g @openscout/scoutIn a terminal, replace the path below with the absolute path to your project:
Fills every command on this page that needs it.
cd /absolute/path/to/projectscout setupscout doctorscout runtimesOn first run, setup asks where you keep your projects. Accept the suggested folder or enter the folder containing your project, such as ~/code. Setup starts the local broker that routes the review request. Before proceeding, confirm that doctor reports a reachable broker and scout runtimes lists Codex with State: ready. Resolve any missing authentication or setup reported there.
More setup help: quickstart and plugins
For installation details, use the quickstart. For commands inside your existing agent chat, follow the Claude Code setup guide or Codex setup guide. The CLI commands below work without those plugins.
2. Make one small change in Claude Code
Open your project in Claude Code and ask it to implement a small, specific change. For example, in an app with a search form:
Prevent an empty search from being submitted. Keep the change focused, run the relevant checks, and leave the changes uncommitted for review.
Use a task that fits your project. Once Claude finishes, inspect the changed files from the same project directory:
git status --shortgit diffgit diff --cachedKeep the checkout stable while Codex reviews it. The review prompt below asks Codex to inspect staged and unstaged changes plus new files, so check that the file list matches the work you want reviewed. If your changes are already committed, give the reviewer the specific commit range instead.
3. Ask Codex to review Claude’s changes
Run this in the terminal, replacing the project path and the example requirement with your own:
scout ask --project /absolute/path/to/project --harness codex --notify "Review the uncommitted changes, including staged changes and new files. The goal is to prevent an empty search from being submitted. Report correctness issues with file and line references, or explicitly say no findings. Do not edit files or run tests."--project selects the checkout, --harness codex selects the reviewer, and --notify returns after the request is accepted so you can check the result separately. By default, the Codex session runs in the background. Add --foreground if you want it to appear as a task in the Codex app.
Scout returns a receipt containing a ref, a token such as ref:1a2b3c4d. Copy your actual token, including the ref: prefix; the token here is only an example. That reference is your handle for this review; an accepted or queued receipt is not yet a completed review.
asked … · flight … · ref:1a2b3c4d. Scout will surface the completion when it arrives.4. Read the findings
Replace RETURNED_REF with the reference from your receipt:
Paste the ref, or the whole receipt line.
scout wait RETURNED_REF --timeout 120Look for State: completed followed by Output: with Codex’s reply. A useful finding names the file and line, explains what could go wrong, and gives the input or condition that triggers it. Codex may also report no findings.
Invocation: …
State: completed
…
Output:
Codex’s review appears hereStill running, or it failed?
If the wait times out, run the same wait command again. A timeout does not cancel the review. To inspect its current state, use:
scout status RETURNED_REF --jsonIf Scout reports a failure, read the error and address the reported cause, such as missing authentication. While work is pending, keep checking the same reference so you do not accidentally launch another review.
5. Follow up using the same reference
Ask Codex to explain a finding in the existing review conversation:
scout ask --ref RETURNED_REF --notify "Explain the highest-priority finding and give a concrete input that triggers it. If there were no findings, describe the main edge cases you checked. Do not edit files or run tests."Using --ref continues the routed reviewer’s context. Scout prints a new receipt for this follow-up. Replace FOLLOWUP_REF below with that new reference to read the answer:
The ref from the follow-up receipt.
scout wait FOLLOWUP_REF --timeout 120Keep using --ref for questions about this review. A new --project … --harness codex request starts a fresh reviewer session.
Evaluate the finding before acting on it. For example, if Codex identifies whitespace-only searches as an overlooked case, ask Claude Code in your original implementation chat to check that case, fix it if needed, and run the relevant tests. A review with no findings is useful evidence, but it does not replace testing.
Finish line
You have completed the loop when you have a review result, a follow-up in the same reviewer conversation, and a decision about which findings to apply. For a worked example of reproducing a finding and checking its correction, continue with the Claude Code and Codex code review guide.
Do you need Scout to use them together?
For an occasional review, you can open both tools in the same Git checkout. Let one finish its change, then give the other the requirement and ask it to review the diff without editing. Keep the files stable during the review, and bring any useful findings back to the implementing conversation. Sharing files does not share conversation history, so include the decisions the reviewer needs.
Scout helps when you want to request that review from a command or connected agent, check its status, and return to the same reviewer conversation with a saved reference. You still decide which findings to apply and run the relevant tests. The steps below show that tracked workflow from setup through follow-up.
If Scout is already running and you only need the review commands and an illustrative finding, use the code review guide.
Can Claude Code review Codex’s work instead?
Yes. Make the change in Codex and select --harness claude when requesting the review. The wait and --ref follow-up steps stay the same. See the Codex setup guide for that direction, the Claude Code setup guide to start from Claude, and all integrations for other supported clients.
Prepare the review brief
Use the agent handoff builder to make the goal, inspected revision, scope, and return format explicit. Read how to hand off context before adding another reviewer.