Claude Code /code-review Ultra Uses Multi-Agent Cloud Review

Anthropic has quietly added a heavier gear to Claude Code's review toolkit. Type /code-review ultra and instead of the usual local pass, Claude Code ships your branch off to a remote sandbox where a whole fleet of reviewer agents goes to work in parallel. The pitch is simple: more eyes, more coverage, and every finding gets independently reproduced and verified before it's reported back, so you're not wading through style nitpicks to find the one real bug.

The tradeoff is that this only runs on Claude Code's web infrastructure, which means it needs a claude.ai login rather than just an API key, and it won't work at all on Amazon Bedrock, Google Cloud's Agent Platform, or Microsoft Foundry, nor for organizations that have Zero Data Retention switched on. In those cases, the command quietly falls back to a local review instead of failing outright.

From the CLI, running the command bare compares your current branch against the default branch, uncommitted changes included, and bundles the whole thing up for upload. Point it at a different base branch, or hand it a GitHub PR number, and it adjusts accordingly — pull requests get cloned straight from the host rather than pulled from your local tree. Newer versions, starting at v2.1.212, added the ability to fetch a base branch that only exists on origin and to suggest the closest branch name if you fat-finger a typo; earlier versions just threw a not-a-branch error.

None of this is free once you're past the initial allowance. Pro and Max accounts get three ultrareview runs as a one-time allotment that doesn't refresh, and after that each review draws from usage credits, typically landing somewhere between $5 and $25 depending on how big the change is — the exact estimate shows up in a confirmation dialog before you commit. A run counts against your allotment the moment the cloud session starts, so stopping one early or watching it fail still burns a free run, though a paid review only bills for the portion that actually executed.

Reviews typically take five to ten minutes and run entirely in the background, so your terminal is free to do other things, or you can just close it and check back later with /tasks. There's also a non-interactive claude ultrareview subcommand meant for CI pipelines, which blocks until the review finishes and prints findings to stdout with proper exit codes. Anthropic frames the whole feature as the deeper option to reach for before merging something substantial, sitting above the lighter /code-review and /review commands meant for faster, everyday feedback.