mrge – AI Code Review That Understands Your Whole Codebase
mrge clones your entire codebase into an ephemeral sandbox, letting its AI reviewer navigate like a senior developer — jumping to definitions, finding.
TL;DR
TL;DR: mrge spins up a cloud sandbox with full codebase context, giving AI reviewers the ability to navigate code like senior developers — jumping to definitions, finding references, and leaving comments directly on changed lines.
Source and Accuracy Notes
- Product: https://mrge.io
- HN Launch: https://news.ycombinator.com/item?id=43692476
- Early users: Better Auth, Cal.com, n8n
- Current pricing: Free while in early beta
What Is mrge?
mrge is an AI-powered code review platform built by Allis and Paul — two engineers who found code review becoming their biggest bottleneck as their team started using AI to write more code.
The problem they identified: AI-written code looks syntactically correct but hides subtle bugs. Human reviewers, buried under a growing pile of PRs, increasingly found themselves rubber-stamping changes without deeply understanding them.
mrge’s approach is fundamentally different from most AI review tools. Instead of sending a diff to an LLM and hoping for the best, mrge clones your repository into an ephemeral cloud sandbox and gives the AI full context — including the ability to navigate the codebase the same way a developer would.
How It Works
- Connect your repo via the GitHub app (two clicks). GitLab support is on the roadmap.
- AI Review kicks in when you open a PR. The AI reviews your changes in an ephemeral, secure container with full codebase context — not just the diff, but the entire project.
- Comments appear directly on changed lines — the AI navigates code using standard tools like “go to definition” and “find references” via an LSP server.
- Sandbox is torn down after review — code is deleted immediately, not stored.
- Human workflow in the web app — files grouped logically (not alphabetically), key diffs highlighted and visualized for faster human review.
Setup Workflow
mrge is a hosted service with a GitHub App integration. There’s no self-hosted option yet, which means your code leaves your infrastructure during the review.
Step 1: Install the GitHub App
Visit https://mrge.io and sign in with GitHub. Install the mrge GitHub App on the repositories you want to cover. The app requires the following permissions:
- Read access to repository contents, pull requests, and commit statuses
- Write access to pull request comments (for leaving review feedback)
Step 2: Open a PR
Once the app is installed, simply open a pull request as usual. mrge automatically picks it up and starts a review.
Step 3: Wait for the AI Review
The AI review runs in mrge’s cloud infrastructure. You can watch progress in the mrge web app — it shows the sandbox being provisioned, the LSP server starting, and the AI navigating your codebase.
When the review is complete, you’ll see inline comments on the relevant lines of your PR, exactly like a human reviewer would leave them.
Step 4: Human Review
After the AI review, open the mrge web app to see the full picture. Changes are grouped by logical connection rather than alphabetically, making it easier to understand the intended impact of a change.
Important diffs are highlighted. You can also add your own comments and approve/request changes.
Step 5: Sandbox Teardown
Once the review cycle is complete and the PR is merged (or closed), mrge tears down the sandbox. Your code is deleted from mrge’s infrastructure.
Deeper Analysis
Cloud-Only Architecture
mrge’s cloud approach is a deliberate choice, not a limitation they’re working around. They argue it lets them run state-of-the-art AI models without requiring local GPU setups, and provides consistent single-review-per-PR pricing for the entire team.
The tradeoff is clear: your code travels to mrge’s servers during the review window. For open-source projects, this is probably fine. For enterprise projects with strict compliance requirements, this is a blocker.
The sandbox model mitigates long-term risk — code isn’t stored after the review. But the transient exposure during review is real.
Desktop App
mrge also ships a desktop app for Mac and Windows. The desktop app provides a snappier interface with keyboard shortcuts, making it feel more like a native tool than a web app. This is worth installing if you do a lot of code review.
Comparison to Existing Tools
Most AI code review tools (like CodeRabbit, Codium, ReviewNB) work by sending the diff to an LLM API. They get the changed lines but lack deep codebase context. mrge’s approach of cloning the entire repo into a sandbox and running an LSP server is meaningfully different — the AI can actually navigate code the way a developer would.
The file-grouping logic (grouping by logical connection rather than alphabetically) is a small but significant UX improvement. Anyone who’s done a lot of code review knows the pain of jumping between unrelated files trying to understand a feature.
What’s Missing
- No self-hosted option — dealbreaker for some teams
- GitLab not supported — only GitHub for now
- No team pricing publicly listed — “free while in early beta” suggests pricing TBD
- Privacy concerns for sensitive codebases — code leaves your infrastructure
Practical Evaluation Checklist
- Connect GitHub repo in two clicks
- AI review fires on PR open automatically
- AI can navigate entire codebase (not just the diff)
- Inline comments appear on changed lines
- Files grouped logically (not alphabetically)
- Sandbox teardown confirmed after review
- Desktop app available (Mac/Windows)
- GitLab support missing (roadmap)
- No self-hosted option available
Security Notes
mrge’s security model has a few key aspects:
- Ephemeral sandbox — code exists in the sandbox only during active review, then is deleted
- No permanent storage — mrge explicitly states they don’t store code after the review completes
- Cloud-only — there’s no way to run mrge on-premises; code always leaves your infrastructure
- Early-stage startup — mrge is still in beta; their security posture may evolve as they scale
For open-source projects, the current model is reasonable. For proprietary code with strict compliance requirements (SOC 2, HIPAA, etc.), evaluate carefully whether the transient cloud exposure is acceptable.
FAQ
Q: Does mrge store my code after the review? A: No. mrge explicitly states the sandbox is torn down and code is deleted after the review completes.
Q: Can I self-host mrge? A: No. mrge is a cloud-only service. There’s no self-hosted or on-premises option yet.
Q: Does mrge support GitLab? A: Not yet. GitHub is the only supported VCS right now. GitLab support is on their roadmap.
Q: How is this different from CodeRabbit or other AI review tools? A: Most AI review tools send the diff to an LLM API and get back comments. mrge gives the AI full codebase context in a sandbox with LSP access, so it can navigate code like a human developer — finding definitions, references, and patterns across the entire repo.
Q: Is there a pricing model? A: mrge is currently free during the early beta period. The plan is to charge for closed-source projects on a per-seat basis, with free access continuing for open-source projects.
Conclusion
mrge’s cloud sandbox approach to AI code review is genuinely different from the diff-sending tools that dominate this space. The ability for the AI to navigate a codebase like a developer — using the same LSP tools, jumping to definitions, finding references — means the review can actually catch patterns that simple diff analysis misses.
The cloud-only architecture is a real constraint for security-sensitive teams, but for open-source projects and teams comfortable with transient cloud processing, mrge is worth trying. Early users include Better Auth, Cal.com, and n8n — teams that handle significant PR volume and have already evaluated the landscape.
The free pricing during beta makes this a low-friction experiment. If you’re doing code review manually or using a tool that only sees the diff, mrge is worth 30 minutes of your time to see what a full-context AI reviewer catches that others miss.