Configure
Context connectors
Bring tickets, docs, designs, production errors, incidents, CI failures and customer reports into reviews from issue trackers, docs, Figma, observability, CI and support tools.
Context connectors tell each review what a change is meant to do, where production is failing, why CI failed and what customers report. Connect them in the console under Integrations, which groups them into the categories below. You need the settings.manage permission.
Only what is linked, pinned or mapped is read. MergeHarbor reads the tickets, documents, designs and support tickets a pull request (or its linked ticket) points to, documents an admin pins for a repository, the log groups, services and support tags an admin maps to a repository, and the failed CI jobs of the pull request’s head commit. It never crawls, lists or searches your workspace to find more.
What connectors add to a review
- Acceptance criteria checked. For a linked ticket, each acceptance criterion is marked Met, Not met or Unclear in the review summary, with
file:lineevidence. This is part of the review itself, not an extra model call. - Docs as review context. Linked or pinned documents are given to the review as the specification for the change, and comments that relied on one link to it.
- Production errors flagged. When a change touches files in the stack frames of recent production errors, the review is told and the summary says so.
- CI failures explained. Failing jobs on the head commit are tied to the lines that caused them, and PR chat can answer “Why did CI fail?”.
- Customer impact. Linked or tagged support tickets add what customers report about the changed area, with personal data removed.
- Optional status comment. For Jira, Azure Boards, Linear, Asana, ClickUp and Trello you can turn on write-back: MergeHarbor keeps one comment on each linked ticket, updated after every review, with the findings, confidence and a link. It is off by default.
Docs and knowledge
| Connector | What it reads |
|---|---|
| Notion | Pages linked from the pull request or its tickets, or named in quotes |
| Confluence | Pages linked from Jira Cloud issues, through the Jira and Confluence connection |
| Google Drive | Docs, Sheets, Slides and PDFs linked from the pull request or its tickets, or pinned |
| OneDrive and SharePoint | Word documents, PDFs and other files behind sharing links, or pinned |
Issue tracking
| Connector | What it reads | Write-back |
|---|---|---|
| Jira | Issues named by key (PAY-12) or link; Jira Cloud or Data Center |
Optional |
| Azure Boards | Work items named AB#123 or by link, and work items linked to Azure DevOps pull requests; Services or Server |
Optional |
| Linear | Issues named by key or link | Optional |
| Asana | Tasks linked from the pull request, its commits or its branch name | Optional |
| ClickUp | Tasks linked from the pull request or its commits, or named CU-<id> in the branch |
Optional |
| Trello | Cards linked from the pull request or its commits | Optional |
MergeHarbor finds ticket keys in the branch name, pull request title and description, and commit messages. Up to three tickets are used per pull request by default.
Observability
| Connector | What it reads |
|---|---|
| Sentry | Top unresolved issues from the last 14 days, with the in-app frames of their latest event |
| Datadog | Error Tracking issues, recent APM error spans, and monitors in alert for the repository’s services |
| PagerDuty | Open incidents on the services mapped to the repository, and incidents that started after a merge |
| AWS CloudWatch Logs | The last day’s error events from the log groups you map to the repository |
| Google Cloud Logging | The last day’s error entries from the services or logs you map to the repository |
They are read in the background at most once an hour (PagerDuty at most every 15 minutes), and only when reviews need them. Reviews never wait for them.
Design
| Connector | What it reads |
|---|---|
| Figma | Frames linked from the pull request or its tickets, as text: layer names, copy, open comments and Dev Mode notes. No images |
CI and builds
| Connector | What it reads |
|---|---|
| GitHub Actions | Failed jobs on the head commit, through the GitHub App |
| GitLab CI | Failed jobs on the head commit, through your GitLab connection |
| CircleCI, Jenkins, Buildkite | Failed jobs they report on the head commit |
MergeHarbor reads the end of each failed job’s log (64 KB, at most 3 jobs per review) only when a review runs or someone asks about CI in PR chat.
Customer support
| Connector | What it reads |
|---|---|
| Zendesk | Tickets linked from the pull request or its tickets, and the 5 latest open tickets with tags you map |
| Intercom | Conversations linked from the pull request or its tickets, and the 5 latest open conversations with tags you map |
Requester and contact fields are never read, and personal data is removed before anything is stored or sent to a model.
Limits and cost
- Ticket and document text adds at most about 3,000 tokens to a review by default, and nothing when nothing is linked.
- At most two documents are read per pull request and per ticket by default, each trimmed to about 8,000 characters before the overall limit applies.
- If a connector fails or is slow, the review continues without it and the connector shows an error in the console.
Data handling
- Least privilege. Each connector asks only for read access, plus comment permission when you turn on write-back. Google Drive uses
drive.fileby default, so MergeHarbor can open only the files you pick for it. CloudWatch uses a read-only role you create, with an external ID unique to your organization. - Redaction. Text that looks like a secret is redacted before it is cached or sent to a model, and ticket, page and error text is never written to logs. Support tickets and cloud log errors also have personal data removed.
- Short-lived cache. Ticket and document text is cached for a short time (10 minutes by default) and deleted within 24 hours. Linked tickets on a pull request keep only their key, link, status and the requirements check. Only if an administrator turns on Keep linked documents in the knowledge base are documents read for issue triage kept longer: for the number of days they choose (1–365) after last use, and deleted at once when the setting is turned off.
- Encrypted tokens. Tokens and keys are encrypted at rest, bound to their connection, and never shown again.
- Scoped per repository. You can limit each connector to some repositories; reviews of other repositories never read it.
- Disconnect. Disconnecting revokes the token where the service allows it, then deletes tokens, pins and cached text.
PostHog is on the roadmap. Choose Request this on the Integrations page to ask for one, or Missing a service? Suggest one for anything else.