How PR data flows
When a pull request triggers a review, the PR diff is fetched by the ReviewAgent Lambda function running in MergeWatch’s AWS account (us-east-1). The diff is held in-memory for the duration of the review and is not persisted to disk, S3, or any database after the review completes.
Bedrock model calls
All LLM inference runs on Amazon Bedrock within MergeWatch’s AWS account. The primary model isus.anthropic.claude-sonnet-4-6 for review agents. Lighter tasks such as summary generation and diagram creation use us.anthropic.claude-haiku-4-5-20251001-v1:0 for cost efficiency. Model inputs (diff content, agent prompts) and outputs (review findings) exist only for the duration of the API call and are not logged or stored by MergeWatch beyond the review lifecycle.
What is stored
Review metadata is written to DynamoDB in MergeWatch’s AWS account. This includes:- Timestamps (review started, completed)
- Repository name and PR number
- Merge-readiness score
- Per-agent latency data
- Settings snapshot (configuration used for the review)
- GitHub reactions data
- User feedback (thumbs up/down)
- Review status (success, error, skipped)
Credential storage
GitHub App credentials (private key, webhook secret, app ID) are stored as SecureString parameters in AWS SSM Parameter Store within MergeWatch’s AWS account, encrypted with AWS KMS.Configuration file
Your.mergewatch.yml configuration file stays in your GitHub repository. MergeWatch reads it at review time via the GitHub API but never copies or stores it.
What MergeWatch sees vs. does not see
Maximum data isolation
For environments where no code may leave your infrastructure — even transiently — use the self-hosted deployment. In self-hosted mode, all processing happens on your own infrastructure with your own LLM provider. No data is sent to MergeWatch.Self-Hosting Overview
Deploy MergeWatch on your own infrastructure for full data isolation.
Architecture
See how SaaS and self-hosted architectures compare side by side.