MergeMind: AI PR Reviewer for Azure DevOps

Automated AI code review for Azure DevOps pull requests. When a PR is assigned to a configured reviewer identity, this service fetches the changes, runs an AI review, and posts inline comments + a sticky summary back to the PR. Supports incremental rounds when developers push new commits.

What you get

• 🔁 Multi-round reviews — Round 2+ automatically classifies prior findings as still-open, resolved, or regressed; sticky summary updates in place.
• 🛑 Cost guardrails — daily review/token caps, per-PR rounds cap, per-review token budget with graceful degradation (full → no_related → diff_only → refused).
• 🔌 Multi-provider AI — OpenAI, Azure OpenAI (with Managed Identity), or Anthropic (Claude) — all behind a single ResponsesClient interface. Configurable via env.
• 🧠 Multi-language context — TypeScript, JavaScript, Python, C#, Java dependency analysis pulls in related files (imports, callers) so the AI sees more than just the diff.
• 🚫 Won't-fix marking — mark individual findings as "won't fix" from the Web UI; the corresponding ADO thread is closed automatically. Won't-fix findings can be hidden from the round view with a toggle.
• 📬 Admin email alerts — SendGrid notifications for permanent failures (auth, quota, poison jobs) with per-(type, scope) throttling.
• 👀 Built-in web UI — read-only React UI at the same port: dashboard with live SSE updates, PR/round detail, retry-failed-posts, health.
• 🪶 Single-process — one Bun process, one SQLite file. No Redis, no queue infrastructure. Reviews run serially to keep cost predictable.

Installation & Running

Using Docker (Recommended)

docker pull your-registry/merge-mind:latest
docker run -p 53847:53847 \
  -v $(pwd)/data:/app/data \
  --env-file .env \
  your-registry/merge-mind:latest

Using Bun

If you have Bun installed:

bun install -g merge-mind
merge-mind

Connect Azure DevOps

In your ADO project → Project Settings → Service Hooks → New Subscription:

1. Service: Web Hooks
2. Trigger: Pull request created (also create a second subscription for Pull request updated — both call the same webhook URL)
3. Filters: leave default (the service does its own filtering by reviewer)
4. Action → URL: https://your-host/webhook/azuredevops (must be HTTPS reachable from ADO)
5. Action → Basic authentication:
   • Username: matches WEBHOOK_USER
   • Password: matches WEBHOOK_PASSWORD
6. Save → Test to verify a 200 comes back.

For the trigger to fire, also assign the configured reviewer identity (the email/group in REVIEWER_TRIGGER_LIST) to the PR. The reviewer can be a real user account or a designated bot account.

Configuration reference

Environment variables

| Var | Required | Default | Description | |---|---|---|---| | WEBHOOK_USER | no | ai-reviewer | Basic Auth username for the ADO webhook | | WEBHOOK_PASSWORD | yes | — | Basic Auth password for the ADO webhook | | AZURE_DEVOPS_ORG_URL | yes | — | e.g. https://dev.azure.com/yourorg | | AZURE_DEVOPS_PAT | yes | — | PAT for ADO API access | | REVIEWER_TRIGGER_LIST | yes | — | Comma-separated reviewer identities; PR triggers when any match | | AI_PROVIDER | yes | — | openai | azure-openai | anthropic | | OPENAI_API_KEY | if openai | — | | | OPENAI_BASE_URL | optional | — | For self-hosted OpenAI-compatible endpoints (vLLM, Ollama, etc.) | | AZURE_OPENAI_ENDPOINT | if azure-openai | — | e.g. https://yourname.openai.azure.com | | AZURE_OPENAI_API_KEY | optional | — | If unset, falls back to DefaultAzureCredential (Managed Identity) | | ANTHROPIC_API_KEY | if anthropic | — | | | AI_MODEL | yes | — | Model id, e.g. gpt-5, claude-sonnet-4-6 | | SENDGRID_API_KEY | optional | — | Enables admin email alerts | | NOTIFY_ADMIN_RECIPIENTS | optional | — | Comma-separated email addresses for alerts | | DB_PATH | optional | data/ai-reviewer.db | SQLite file path | | PORT | optional | 53847 | HTTP port | | LOG_LEVEL | optional | info | trace | debug | info | warn | error | fatal |

Per-repo .ai-reviewer.yml

Drop a file at the root of any repo you want reviewed to override defaults for that repo.

enabled: true                      # set false to disable AI review for this repo
exclude_paths:
  - "dist/**"
  - "build/**"
custom_prompt_addendum: |
  This repo uses TypeScript strict mode. Flag any `any` usage in new code.
severity_thresholds:
  post_inline:  ["must-fix", "should-fix"]
  post_summary: ["must-fix", "should-fix", "nit"]
limits:
  max_changed_files: 50
  max_tokens_per_round: 200000

Web UI

Visit http://localhost:53847 to access the dashboard.

| Route | What it shows | |---|---| | / Dashboard | Live queue status and recent activity. | | /prs | Searchable list of all reviewed PRs. | | /rounds/:id | Full report markdown, every finding grouped by severity. | | /failed-posts | Findings whose ADO comment posting failed. | | /health | Connectivity and usage status. | | /settings | Runtime configuration (models, reasoning, etc). |

Operations

Health check

curl -u $WEB_UI_USER:$WEB_UI_PASSWORD http://localhost:53847/api/health

Logs

The service logs structured JSON.

Backup

The DB is a single SQLite file at data/ai-reviewer.db.

License

MIT