serge reviews pull requests with an OpenAI-compatible LLM and posts
inline comments through GitHub’s Pull Request Reviews API. The default persona
is Serge, and the default trigger is @askserge.
The reviewer reads the PR diff, applies repository-specific rules, optionally uses read-only repository context tools, validates every proposed inline comment against the diff, and publishes only comments that point to real diff positions.
serge also has a write-capable side. The tasks flow lets CI hand serge a failure report and get back a pull request with a proposed fix — the same trust pattern as reviews (the LLM only proposes; serge applies and publishes). It is off by default and gated behind GitHub Actions OIDC.
What serge does
| Capability | What it produces | Guide |
|---|---|---|
| Review | Inline PR comments validated against the diff | How it works |
| Fix (write-capable) | A fix PR or follow-up commit from a CI failure report | Tasks flow |
Choose a Mode
| Mode | Use it when | Where it runs |
|---|---|---|
| GitHub Action | You want per-repo control via a workflow file | GitHub Actions |
| GitHub App webhook | You want a hosted reviewer across many repos with no per-repo workflow | Your server |
| Web app | You want to edit or discard LLM output before it reaches a PR | Your server |
Start with Getting started, then use the guide for the deployment mode that fits your repository. New to the UI? Take the web app tour.
Core Features
- OpenAI-compatible chat completion endpoints.
- Inline PR comments validated against actual diff lines.
- Trigger comments and follow-up replies on inline review comments.
- Repository rules from
.ai/review-rules.md. - Optional repository context from
.ai/context-script. - Optional helper tools from
.ai/review-tools.json. - Human-in-the-loop staged reviews through the web app.
- Write-capable tasks flow: CI sends a failure report, serge opens a fix PR (off by default, OIDC-gated).