The product

A Review is one AI code review of one pull request at one point in time. When it completes it has a Verdict, APPROVED or NEEDS_REVISION, plus Issues (each with a severity, a recommendation and optionally a file and line) and PR-level notes on breaking changes, test coverage, dependencies or docs.

Reviews run on GitHub and Azure DevOps pull requests. A dry run does the full review and posts nothing back to the PR.

Developer reactions to the threads PR Bot posted are recorded as learning signals and classified as fix, exception, dismiss, agreement or noise. Written replies wait in an approval queue until a person approves or rejects them.

Approved signals become repo context: guidance for that repo that goes into every later review of it.

Monitored PRs track a specific Azure DevOps PR against an approval threshold, with MS Teams alerts. Build monitors show recent build runs as health dots. Release monitors analyse failed Playwright test runs in Azure DevOps Classic Releases and deliver the result as a Teams card.

PR Bot does not call a model directly. Every review, classification and analysis goes through the MCP bridge, a separate mcp-agent-bridge service you run, with claude, codex or copilot as the MCP provider.

The web app (Next.js) is where your team works. The API (Hono, OpenAPI) and a Temporal worker do the rest. A CLI, an MCP server and an agent skill expose the same four core operations: trigger a review, get review status, list bypassed reviews and get PR stats.

The desktop app (Tauri) bundles the web app, API, worker, a Node runtime and Temporal, with a local SQLite database.