Skip to Content
OpenInspect

OpenInspect documentation

What OpenInspect does, how a session runs from prompt to pull request, and where each part of the product is documented.

Last reviewed View as MarkdownEdit on GitHubGive feedback

OpenInspect is a background coding agent. You describe a piece of work, the agent runs it in an isolated cloud sandbox with a full development environment, and you come back to review the result: a response, a diff, and usually a pull request.

The OpenInspect home screen with the session inbox on the left and the new-session composer in the middle
The home screen: inbox on the left, composer with repository, branch, model, and skills pickers in the middle.

The agent does not need you to stay connected. Send a prompt from the web app, Slack, a GitHub pull request, a Linear issue, or an automation, close the tab, and check the session later. Several people can watch and prompt the same session at the same time. Prompts record who sent them; commits use that person's source-control identity when available, or the configured Git user (the signing account when signing is enabled) otherwise.

Start here

One lifecycle

Every session, whether it starts from the web composer, a Slack message, or an automation, follows the same path.

From your side, each session comes down to four actions:

  1. Choose a target: a single repository and branch, an ad-hoc set of up to 10 repositories, a saved environment, or no repository at all.
  2. Describe the work in one prompt: the goal, the constraints, and how the agent should prove it is done. Images can be attached.
  3. Review the timeline, the final response, and Changes (the diff compared with session start), plus the tests or screenshots the agent produced.
  4. Iterate: send another prompt on the same branch, ask for a pull request, or open the sandbox terminal and code server and finish the work yourself.

Capability map

AreaWhat it coversStart with
SessionsThe session page, follow-ups and stopping, reviewing diffs, sandbox tools, sharing, child sessions, statuses, and recovery.Session page
PromptingWhen to delegate, how to write a prompt the agent can finish, templates, and workflow recipes.When to delegate
ConfigurationLifecycle scripts, environments, prebuilt images, secrets, sandbox settings, source control, skills, MCP servers.Lifecycle scripts
ModelsChoosing a model and reasoning effort, the two agent harnesses, and provider accounts.Choosing a model
AutomationsSessions started by a schedule, an inbound webhook, a GitHub event, a Slack message, or a Sentry alert.Automations
IntegrationsSlack, GitHub (reviews, mentions, PR Feedback Autofix), Linear, personal access tokens, and the MCP server.Integrations
AdministrationWorkspace access and roles, Teams, the audit log, analytics, data controls, security, and deployment.Workspace access

Where people stay in control

The built-in pull-request tool opens PRs but does not merge them. The agent can also run the GitHub CLI, including merge commands, when its credentials and repository rules permit it. GitHub PRs are authored as the prompting user when a token can be resolved, or as the App otherwise; an automation run's own PRs are always authored as the App. Enforce branch protection, required reviews, and status checks in your repository.

Sessions have Workspace, Team, or Private visibility, separate from their fixed ownership. Team-owned actions require current team membership even for administrators, except for reads and collaborator self-removal; private sessions support explicit collaborators and audited Owner break-glass reads. Non-private Team read isolation requires enforcement to be on, rather than the default shadow; see Deployment. Every prompt is attributed to its author, and administrators can read the audit log to see who did what. See Multiplayer and sharing for the access rules.

On this page