Korbit

You will like Korbit for two chores it takes off your plate: it writes the pull request description, and when a comment is wrong you argue with the bot in the thread instead of dismissing it, which turns review noise into a conversation. That is the daily trait, and on a team with a description-writing problem it is worth the seat on its own.

It will not run your tests or touch the code. Pick it for a team drowning in undocumented PRs and tired of reviewers asking what a change is for. Pick CodeRabbit if you want the crowd's default and the largest install base.

The risk is throughput. The FAQ states a maximum of two parallel scans with automatic queuing, and the Max tier sells higher concurrency limits, which confirms the ceiling is real rather than a doc artifact. Twenty pull requests at once get reviewed serially, and the twentieth waits behind the other nineteen.

The consequence is that a Friday merge train waits in line behind itself, and a required-review rule on the bot turns that queue into a release delay. Published concurrency per tier would change this verdict. What it does right: Essential and Comprehensive review modes let you throttle the noise before it reaches the thread.

The documented mechanism is 1. detection of functional, performance and security issues with configurable categories and severities, and 2. adaptive learning from user feedback: ignored issues, resolutions, thumbs up and down. A feedback loop is a claim of improvement over time, and no benchmark, precision or before-and-after number is published to show it converges.

The consequence is that a team cannot tell whether dismissing comments trains the reviewer or merely hides them, and the two look identical from the thread. A precision curve over the first ninety days would settle it. Asserted, not measured.

Korbit Technologies picked a different buyer. The console and its insights reports, with Slack digests at the higher tier, are aimed at the engineering manager, and managers renew things that produce charts, so the retention hook is the dashboard rather than the review. Max free for open-source repos is the distribution play everyone runs.

Moat: thin, the reports are copyable and the reviewer is a model call. Likely acquirer: Atlassian, which needs a reviewer for Bitbucket and already sells to the same manager. Position: small, and sized by how many managers want a chart more than engineers want a bot.

Procurement: Pro is $12 a user on annual, so $720 a month for sixty; Max is $18, so $1,080, and adds custom policies. The FAQ states code is retained only for the duration of the scan, with zero-day retention at the model providers, which answers the questionnaire's hardest question before it is asked.

On-prem is Enterprise. SSO is not listed on the pricing page, which for a Git app installed at the organization level is the gap. CI fit is native, it runs on every pull request. Onboarding is a Git app install. Approved with conditions: SSO confirmed in writing.

Nothing to read, nothing to run, nothing to key. OpenAI and Anthropic are picked for me, there is no bring-your-own model, no local option, no MCP client or server, and no command line, so there is no script I can write against it and no config file I can version.

Custom policies for enforcing coding rules are the one configuration surface, and the Bitbucket install is workspace-level access, which is a lot of permission for a tool I cannot inspect. Nothing to fork. If the reviewer is good, it is good on their terms, and I have no way to check. Not mine.