Skip to content

DiffHarbor

Self-hosted AI code review for GitHub and GitLab. It runs the linters and tests your repo already has, verifies every finding against the code, and comments only what holds up.

A small Go server takes the webhook, runs your checks on the changed lines, lets an agent (Pi) review the diff with read, grep and bash in a checkout, has a second model verify every finding against the code, and posts what survives as inline comments with committable suggestions.

Verified findings

A second, cheaper model checks each finding in the code and drops what it can’t back up.

Your own checks as context

gitleaks, tsc, biome, eslint, oxlint, gofmt, vet, staticcheck, golangci-lint, deadcode, fallow, squawk, shellcheck, luacheck, actionlint, zizmor, osv-scanner and the tests of changed packages, on the changed lines.

Incremental

A push reviews only files whose changed lines changed. A rebase or a merge of main reviews nothing.

Committable fixes

A local fix becomes a suggestion block; fixed findings get their threads resolved.

Chat and learnings

Reply to a finding or mention the bot; corrections become rules for the next reviews.

Any model Pi supports

OpenAI, Anthropic, DeepSeek, AWS Bedrock, Azure OpenAI or a gateway in front of them, with automatic fallback.

PR reports

An optional HTML page per PR with a reading order, diffs with findings, call graphs and merge risks.

Small and self-hosted

One container, Go standard library only. Your code stays on your server and with your model provider.