Introduction ​
CodeLotse reviews pull requests, triages issues, suggests reviewers and turns issues into draft pull requests. It runs next to your code host, uses the language model provider you choose with your own key, and posts its results as comments, labels and statuses where your team already works. Everything it writes is a recommendation: your team decides what gets merged.
A Lotse is a harbour pilot. The pilot comes aboard, knows the waters and keeps the ship on course, while the crew stays in charge.
What CodeLotse does ​
| Feature | What you get | Page |
|---|---|---|
| Pull request review | A summary, inline comments with suggested changes and a status on the head commit. New commits are reviewed incrementally. | Pull request reviews |
| Approvals | Outputs wait for a person to post or reject them, with a preview of exactly what will be posted. | Approvals |
| Commands in comments | explain, why, fix, ignore, re-review and more, by mentioning the bot. | Commands in comments |
| Issue triage and replies | Labels from an allowlist, possible duplicates, requests for missing information and grounded replies. | Issue triage and replies |
| Issue to pull request | A maintainer asks for a fix; an agent works in a sandbox, runs your tests and opens a draft pull request. | Issue to pull request |
| Checks | Deterministic scanners next to the review, and a check of the acceptance criteria of the linked issue. | Checks and acceptance criteria |
| CI failure analysis | An explanation of why a check of a pull request failed. | CI failure analysis |
| Rules and learnings | Team rules, rule files such as AGENTS.md, learnings and suppressions of noisy findings. | Rules, suppressions and learnings |
| Reviewer suggestions | Reviewers and assignees based on code ownership and recent work, with strict privacy controls. | Reviewer suggestions |
| Analytics | Team-level outcomes and model costs. Never a ranking of people. | Analytics |
What CodeLotse does not do ​
- It never merges, approves or closes pull requests, and it never closes issues.
- It does not post anything until an administrator turns a feature on. Setup recommends Shadow mode, which reviews without posting, and you can require approval before anything is posted.
- It never removes labels and never applies protected labels such as
securityorp0; it only proposes them. - The fix agent never pushes to your default branch. It opens draft pull requests from its own branch, and only when a maintainer asks for it.
- Analytics never rank or score people, and reviewer suggestions show no scores.
- It does not train models on your code. Feedback only tunes thresholds and proposes suppressions for your organisation.
Editions ​
| Edition | Who runs it | Where to start |
|---|---|---|
| Self-hosted open core | You, on Docker Compose or Kubernetes, also in air-gapped networks. Licensed under Apache-2.0. | Quickstart (self-hosted) |
| CodeLotse Cloud | Quavon UG (haftungsbeschränkt) runs it for you in an EU data centre at codelotse.com. You still bring your own model key. | Cloud documentation |
Both editions have the same features for reviewing, triage, fixes and assignment. The cloud edition adds sign-up, plans and billing.
Where to go next ​
| You are | Read |
|---|---|
| A developer whose repository uses CodeLotse | How it works, The web app, Pull request reviews, Commands in comments |
| An organisation administrator | Organisation settings, Repositories and policy, Models and keys, Members and roles |
| An operator who runs the instance | Quickstart (self-hosted), Install with Docker Compose, Instance administration, Operations |
Unfamiliar terms are explained in the glossary.