Anthropic officially launched Ultrareview for Claude Code on Wednesday, April 22, 2026. This “deep code analysis” feature represents a major shift from traditional linters, utilizing a parallel fleet of AI agents to autonomously hunt for bugs, logic errors, and security vulnerabilities within a cloud-based sandbox.
Unlike the standard /review command, which provides quick local feedback, Ultrareview is designed for “long-horizon” verification before merging major pull requests.
1. How Ultrareview Works: The “Multi-Agent Fleet”
Ultrareview is powered by a multi-stage orchestration layer that separates the act of finding a bug from the act of verifying it.
- Stage 1: Setup: Provisions a remote cloud sandbox (managed by Anthropic) and clones the relevant branch or GitHub PR.
- Stage 2: Find (The Swarm): Spawns a fleet of 5 to 20 sub-agents (Opus-class for logic, Sonnet-class for style). These agents approach the code from different “personas”—such as a security auditor, a database architect, or a performance engineer.
- Stage 3: Verify (The Secret Weapon): For every potential bug found, a separate agent attempts to independently reproduce the issue. Findings that cannot be verified are discarded, which reportedly keeps the false positive rate below 1%.
- Stage 4: Dedup: Consolidates findings from the fleet into a single, prioritized report.
2. Command Usage and Features
The feature is integrated directly into the Claude Code CLI (v2.1.86+).
| Command | Action |
/ultrareview | Reviews the diff between the current branch and the default branch. |
/ultrareview <PR#> | Fetches and reviews a specific GitHub Pull Request. |
/tasks | Monitors the progress of background reviews (typically takes 5–10 mins). |
/remote-env | Configures the cloud environment where the review agents run. |
3. Pricing and Availability
Anthropic is positioning Ultrareview as a premium “agentic” service rather than a standard chat feature.
- Eligibility: Available for Pro, Max, Team, and Enterprise users.
- Trial Period: Pro and Max subscribers receive three free Ultrareview runs available through May 5, 2026.
- Billing Model: After the trial, Ultrareview will be billed as “extra usage” rather than counting against a user’s standard rolling token window, reflecting the high computational cost of running parallel agent fleets.
4. The “Project Glasswing” Connection
Ultrareview is the first public implementation of the defensive safeguards developed during Project Glasswing, Anthropic’s internal initiative to use AI to find “zero-day” vulnerabilities.
While the most dangerous capabilities of the “Mythos” model remain restricted, Ultrareview uses the same underlying reasoning logic to detect complex flaws—like the 27-year-old OpenBSD bug discovered during internal testing—without exposing the offensive “exploit-building” tools to the public.
