Connect an AI agent to GitHub through a currently supported integration or credential method, then grant only the repository actions required for its task. The safe default is read-only repository permission for viewing code, with edits kept on a separate local branch; grant repository write permission only when the workflow actually needs to push, and add issue or pull-request permissions the same way.
Use current official documentation from both the agent vendor and GitHub because installation and permission options can change. Prefer a supported organization-managed integration and short-lived, narrowly scoped credentials where available. Configure each control at its own layer: GitHub App permissions and repository access for the credential; branch protection and required checks for branch updates; and deployment environments for branch or tag restrictions and environment secrets. A token setting does not replace branch or environment rules. Test the connection on a non-sensitive repository and confirm that forbidden actions fail.
Do not expose deployment keys or production secrets just because the agent can create code. Treat issue comments, pull-request text, and repository files as untrusted input that could influence an automated workflow. Require approval before privileged commands or releases.
A successful connection is one you can explain and disable. After the test, inspect audit records and rotate any temporary credential. Then add one bounded coding task and require a reviewable diff with checks, keeping merge authority with your existing reviewers.
Source: GitHub App best practices; GitHub protected branches; GitHub deployment environments.
