Skip to main content

Copy this into your coding agent

Paste into Cursor, Codex, Claude Code, or another coding agent in your project.
1

Connect GitHub

Choose Connect GitHub and grant access to the repositories you want to test. Bench does not need every repository in your account.
2

Choose a repository and branch

Open Repositories, choose the repository, then select its branch and start a scan. A connection is not the same as a completed scan.
3

Open your systems

After scanning, open AI systems. Use the repository selector to change projects. Your selected repository stays selected as you navigate.
No repository? Choose Add AI system → Upload prompts and add a text, Markdown, JSON or YAML file. Select the model those prompts use. Previously scanned targets appear first, marked Scanned · main, for example. Click the branch to choose another branch or an open pull request. Your saved target stays selected when you return. Scanning dev creates a separate row and systems; it does not replace main. A scan with no detected systems still appears. Use the repository picker to view all branches or one specific target. Systems, findings and evaluation history follow that selection. Choose Scan again to refresh the selected target.

Check a pull request

Choose a pull request from Branch or pull request. Bench scans its head commit and retains it separately as PR #42, for example. Findings and evaluations belong to that PR target, so later source changes do not rewrite older results. This checks the selected source snapshot; it does not post a GitHub Check, merge the PR, or automatically run the entire application. For execution evidence, run real app tests against that same revision. To propose another code change, scan the PR’s source branch first. Listing PRs requires GitHub access to read pull requests; you can still scan an accessible source branch if PR listing is unavailable.

Background refresh

Scanned GitHub repository/branch pairs are checked about every ten minutes. Bench reruns discovery when the source SHA changes and preserves your system identity and manual context. The interval depends on service availability and backlog; failed checks back off before retrying. Refresh happens in the background. Uploaded prompts do not have a GitHub refresh schedule. A scan can remove a source warning when its evidence disappears. It does not prove a production fix or start unrestricted application execution. Automatic real app testing requires a separately enabled real app test setup.

Manage access

Use Manage GitHub access to change the GitHub App’s repository permissions. Removing an AI system in Bench does not delete code or revoke the GitHub App. Historical context can contain evidence from removed components. A restricted API key must cover those repositories too. Legacy evidence with unverified repository history requires a signed-in owner’s review. See framework support for recognition limits and your first bench for the next step.