LiveReview

• Letting low-risk pull requests such as copy changes, dependency bumps or isolated UI tweaks merge without waiting for a senior reviewer
• Ranking the hunks of a large pull request so reviewers read the shared auth or database code before the cosmetic changes
• Running a commit-time review on every developer's machine with git-lrc, with skipped reviews visible in the Git log
• Blocking merges in GitHub Actions, GitLab CI or Azure Pipelines with a team-specific jq rule, such as any security finding or any critical correctness bug
• Checking that engineers understand AI-written changes through summary decks and PR quizzes
• Answering engineering-leadership questions about review adoption, issue hotspots and review cost from Slack or Teams

LiveReview's useful idea is to treat review capacity as the scarce resource and spend it by risk: score every change, let the dull ones through, and rank what remains by how far each hunk reaches. The policy replay and false-approve reporting are the right safeguards for that approach, and they matter more than the headline percentage, because the 30-50% figure is Hexmos's target and the dramatic incident numbers belong to Meta's RADAR work. The incident-trained model is also only as good as the outcome data a team records, so expect the early weeks to rest mostly on the LLM and code-graph signals. Pricing by reviewed LOC with unlimited users is easy to budget and favors large teams with modest change volume, while teams shipping very large diffs should estimate which band they will land in. Buyers who need SOC 2 Type II or an OSI license should look elsewhere; teams that want a reviewer they can self-host and audit across five Git hosts, and that are comfortable with fair-code licensing, should put it on the shortlist.

About LiveReview

Most AI reviewers add comments to every pull request and leave the merge decision to a person. LiveReview, which Hexmos announced as generally available on November 7, 2026, is built around a different question: does a human need to look at this change at all? Each pull request gets one risk score that blends three signals: an LLM's semantic read of the change, an ML model specific to your organization that retrains on the reverts and incidents you record, and a static graph of your codebase that measures how far the change reaches. Inside the diff, every hunk gets a Blast Radius score (how many places call the code, whether it writes to a database, whether callers sit in unrelated parts of the codebase) and a Review Priority score (branching complexity, loop nesting, how much it calls into, whether it has tests), so a three-line change to a shared auth check can outrank a 300-line UI tweak. On top of the score sits auto-approval: you pick a conservative, balanced or aggressive policy, can replay a draft policy against recent pull requests before turning it on, and get a calibration report of what was approved and blocked, the override rate and the false-approve rate. Hexmos says the lowest-risk 30-50% of pull requests can be approved this way; that is a vendor claim, and the 50x-fewer-incidents figures it cites come from Meta's internal RADAR system that inspired the design, not from LiveReview customers. Review can also run earlier than the pull request. git-lrc is a Git hook that reviews staged changes at commit time and records reviews, vouches and skips in the Git log, claude-lrc adds the same review inside Claude Code, and there are pre-push, CI/CD and scheduled default-branch checks, which you turn on per repository. Two features target engineers who approve code they did not really read: a 60-second summary deck for every review and a short quiz generated from the diff. Teams encode standards in a .lrc/ folder of rules in the repository, write merge gates as jq expressions over the review findings and call them from any CI pipeline, and ask an analytics assistant called Livi questions about review data on the web or in Slack, Microsoft Teams or Discord. Pricing follows reviewed lines of code rather than seats: 30,000 reviewed LOC a month free, then fixed bands from $32 for 100,000 LOC to $1,024 for 3.2 million, with unlimited users on paid plans. The code is on GitHub under a modified Sustainable Use License, which is source-available rather than open source, and it self-hosts with Docker; without a purchased licence a self-hosted install runs on the free 30K LOC allowance. The tradeoffs: the product is newly generally available, the incident-trained model only learns from outcomes your team records, SSO, local models and supported self-hosting sit on a sales-quoted Enterprise plan, and Hexmos says it is not SOC 2 Type II certified.