Skip to main content
All notable changes to MergeWatch are documented here. The changelog in the source repository is auto-generated by the release script from conventional commits.

​
[0.5.0] — 2026-08-19

First tagged release since 0.1.0. Four months of work: competitive parity, org-level controls, and a sustained push on review reliability.

​
Highlights

  • Org custom agents — define review agents once in the dashboard and enforce them across every repository, with path/language targeting and an optional blocking mode that fails the check on a critical finding. See Custom agents.
  • MCP server — expose MergeWatch review to MCP-compatible clients for pre-commit review.
  • Structured LLM output — agents now emit findings against an enforced JSON schema on every supported provider, eliminating a class of failure where a malformed response silently discarded findings.
  • Burst resilience — a durable job queue with bounded concurrency (SQS on SaaS, Postgres on self-hosted) paces a flood of pull requests instead of hitting the provider’s rate limit all at once. Rate-limited reviews park and retry rather than failing. Tunable via REVIEW_CONCURRENCY.
  • Merge score in the check title — the GitHub check run leads with the verdict (3/5 — 2 findings), so a passing check no longer hides an advisory result. See Review behavior.
  • False-positive reduction — grounding findings against the diff, a claim-verification pass, cross-agent deduplication, clustering of related concerns, and a configurable confidence floor (minConfidence).
  • Configuration that actually applies — minSeverity, maxTokensPerAgent, and postSummaryOnClean are now read by the pipeline; several were previously documented but unwired.
  • Analytics — windowed rollups, p95 time-to-merge, and per-day counters, with the 500-review ceiling removed.
  • Cost visibility — per-PR and dashboard cost for self-hosted deployments, including custom model pricing for models MergeWatch does not price by default.

​
Notes

  • The default dashboard severity threshold is Low (report every tier). Med and High are explicit opt-ins.
  • Docker images are published per release: ghcr.io/mergewatch/mergewatch:0.5.0 and ghcr.io/mergewatch/mergewatch-dashboard:0.5.0.

​
[0.1.0] — 2026-04-04

Initial release of MergeWatch — AI-powered GitHub PR review with a multi-agent pipeline.

​
Highlights

  • Multi-agent review pipeline with six specialized agents: security, bugs, style, error handling, test coverage, and comment accuracy — plus summary and diagram passes for PR context.
  • Two deployment modes: SaaS (Lambda + DynamoDB + Bedrock) and self-hosted (Docker + Postgres + any LLM).
  • Four LLM providers: Anthropic, Amazon Bedrock, LiteLLM (100+ models), and Ollama (local / air-gapped).
  • Codebase awareness — agents can perform agentic file-context requests beyond the diff hunk, bounded by maxFileRequestRounds and maxContextKB.
  • Next.js dashboard with GitHub OAuth, review history, analytics, and per-repo settings.
  • Smart filtering — skip docs-only / lock-file PRs, exclude patterns, diff filtering, draft PR skip, and label-based ignore rules.
  • Respond mode — reply to @mergewatch mentions in PR comments (review, summary, or free-form respond).
  • Inline thread replies — MergeWatch replies when humans respond inside review-comment threads it started.
  • Inline PR review comments for critical findings, anchored to changed lines only (changed-line enforcement).
  • Docker Compose setup for one-command self-hosted deployment (server + Postgres + dashboard).
  • Prepaid credits billing for SaaS (Stripe) — 5 free reviews per installation, then pay-as-you-go with optional auto-reload.
  • AGPL v3 license with a commercial license available.

​
Known limitations

  • GitHub Enterprise Server not yet supported (GitHub.com Cloud only).
  • No GitLab or Bitbucket support in v0.1 — GitHub only.
For the authoritative changelog, see CHANGELOG.md in the source repository.