Ultrareview: deep multi-agent code review in the cloud before you merge

/code-review in Claude Code gives fast feedback while you code. But before merging a big change that touches payments or authentication, you want a much deeper pass. Ultrareview (/code-review ultra) runs a fleet of reviewer agents in a cloud sandbox on Anthropic's infrastructure, and every finding is independently reproduced and verified before it's reported. It's a research preview.

How to run it

/code-review ultra                    # current branch vs the default branch
/code-review ultra develop            # against another base
/code-review ultra 1234               # review a pull request on github.com
/code-review ultra check my auth changes   # with a plain-language note

• Before launch, a confirmation dialog shows the scope, your remaining free runs and the estimated cost.
• The review runs in the background for about 5–15 minutes; keep working and track it with /tasks.
• When done, findings appear as a notification in your session with file locations and explanations, so you can ask Claude to fix them directly.
• For PRs on github.com, you can choose to post the findings as one comment from your GitHub account (not a review or approval).

Compared with regular /code-review

Requirements and limits

• You need to sign in with a claude.ai account (not just an API key). Pro, Max and Ultra get 3 free runs, one-time, which don't refresh.
• After the free runs, usage credits must be on; a review stopped early still bills for what ran.
• Not available on Bedrock, Google Cloud, Foundry, or organisations with Zero Data Retention, HIPAA or CMEK.
• A branch diff is limited to about 500 files and 8,000 changed lines by default; if it's too big, split it or use PR mode.
• Branch mode uploads your repo state to the cloud; PR mode clones from github.com and uploads nothing from your machine.

Running it in CI

claude ultrareview 1234 --json --timeout 45

The command waits until the review finishes and prints findings to stdout; exit code 0 when complete (with or without findings), 1 on failure or timeout. Note that claude -p '/code-review ultra' exits before the findings arrive.

Risks and caveats

• No findings ≠ safe code: you still need humans to review design and business context.
• Cost scales with size: use it for changes worth it; use /code-review for small ones.
• Closing the session early: if you chose to post to the PR, keep the session open until it finishes, or nothing is posted.

How to apply it

• Dev: Use /code-review while coding and save ultrareview for the last step before opening a big PR.
• Tech Lead: Make ultrareview mandatory for changes touching payments, authentication or customer data.
• PM: Put ultrareview costs in the release budget; $5–25 is far cheaper than a production incident.