Kilo Code Reviewer

Kilo Code Reviewer automatically analyzes pull or merge requests and can also review local code before a developer commits. On repository workflows, the reviewer receives PR/MR metadata, diffs and surrounding file context, then produces inline comments, summary findings, suggested fixes and severity/risk tagging.¹ Users can choose strict, balanced or lenient review styles and focus the agent on security, performance, bugs, code style, tests or documentation.²¹

The product is model-agnostic. Kilo markets access to hundreds of models across the wider platform, and Code Reviewer lets teams select different models depending on depth and cost.²³ This matters because code-review economics can vary dramatically between frontier and budget models; it also avoids turning the reviewer into a wrapper around one fixed model vendor.

Repository policy is one of the stronger design choices. Teams can commit a `REVIEW.md` file to the base branch containing severity calibration, files to skip, verification expectations and sub-agent guidance. Kilo deliberately reads this policy from the base branch rather than the unreviewed feature branch, preventing a PR from rewriting the rules used to judge itself.¹⁴ For larger diffs, the reviewer can use read-only sub-agents and the main reviewer remains responsible for verifying and deduplicating findings before posting.¹

Local review extends the same workflow into VS Code/CLI before a PR exists. `/review` can inspect uncommitted changes, a branch against its base, a commit or a PR.¹ The broader Kilo mobile app now also exposes Code Reviewer management and PR review from iPhone/iPad, including line-level comments, approvals and merge actions.⁵

Platform support is slightly inconsistent across current surfaces. The main Code Reviewer page says GitHub and GitLab are supported and Bitbucket is “coming soon,” while the current iOS app describes Code Reviewer support for GitHub, GitLab and Bitbucket.²⁵ Teams relying on Bitbucket should therefore verify rollout status directly rather than assuming parity from one surface.

Kilo now advertises an unlimited free Code Reviewer path when users select Kilo Auto or a free model, with no credit card required.²⁶ Paid frontier-model reviews consume Kilo model credits; the wider Kilo platform states that model usage is charged at provider rates with zero markup.³ The documentation separately says review compute/time is free during the beta while model reasoning still consumes credits.¹

Kilo publishes concrete vendor-run cost examples from real Hono PRs. In those tests, Claude Opus 4.6 cost $0.73 on a 338-line/9-file PR and $1.34 on a 598-line/5-file PR; Kimi K2.5 cost $0.07 and $0.05 respectively.⁶ These numbers are useful for understanding order of magnitude, but they are Kilo's own tests and should not be treated as independent price/performance validation.

For a 10-person team handling roughly 660 PRs/month, Kilo estimates about $660 using frontier models for every review, $165 with a mixed strategy and $40 with budget models.⁶ Because cost depends on model choice, context size and token use, buyers should model their own repository mix rather than turning those examples into fixed subscription pricing.

Kilo's technical transparency is strong. The Code Reviews documentation describes exactly what data the reviewer receives, how repository guidance works, when sub-agents are used and what hard read-only constraints apply.¹ The open-source Kilo repository is MIT-licensed and has a large public development footprint, although the hosted Code Reviewer service itself should not be assumed to be fully represented by the open-source client code.³

The company also has significant platform scale: Anaconda announced its acquisition of Kilo in July 2026 and described the Kilo platform as used by more than three million developers.⁷ That is evidence of ecosystem adoption, not a claim that three million developers use Code Reviewer specifically.

The decisive current trust issue is the August 2026 Metabase incident. Anaconda disclosed that an unknown actor accessed customer records through Kilo's business-intelligence provider; a subset of Kilo users or organizations may have had partial or full prompts plus data such as names, email addresses, billing addresses and/or location exposed. A small subset of Slackbot users also had Slack access tokens exposed and invalidated.⁸ Payment card information was not believed exposed, and affected users were notified. Anaconda subsequently added a data-export path and reiterated account-deletion options.⁸

That incident does not establish that Code Reviewer itself was exploited or that source repositories were exposed. It does demonstrate that Kilo's broader data/analytics architecture had a material security failure through a third-party component. For teams sending proprietary code context to review infrastructure, this should be part of procurement and incident-response diligence rather than hidden behind a generic “enterprise-grade security” label.

Independent review-quality evidence is the other major gap. Kilo's published Hono tests measure model cost and issues found, and its product testimonials include claims that Kilo found bugs missed by other reviewers.⁶² No current independent benchmark with labeled defects, precision/recall, false-positive rate and multiple competing reviewer products was verified. Model choice also means “Kilo accuracy” is not one fixed quantity.

Kilo Code Reviewer is strongest for teams that want review automation but dislike model lock-in. A team can use a free/budget model for routine PRs and reserve a frontier model for security-sensitive or release branches. Custom focus areas and `REVIEW.md` make it more adaptable to a real codebase than a generic one-prompt reviewer.¹⁶

It is also well suited to developers who want the same review concept before and after push: local `/review` for uncommitted work, then automatic GitHub/GitLab review once the PR opens.¹ The broader Kilo ecosystem—VS Code, JetBrains, CLI, cloud agents and mobile—adds continuity for teams already using Kilo.²⁵

It is a weaker fit for organizations that require a long incident-free security record, independently validated review accuracy, or a single deterministic model configuration across all projects. Security-conscious enterprises should specifically review post-August data flows, retention, analytics tooling and sub-processors before rolling it across sensitive repositories.⁸