# Bluebear for GitHub Copilot

**Roll out GitHub Copilot with visibility and runtime control.** Govern what Copilot's agent mode and CLI run, write, and install on developer machines, track drift in your organization's Copilot policies, and see what the Copilot coding agent does in your repositories. Review the findings together in Bluebear, next to the ones from Claude Code, Cursor, and your other agents.

Available now for GitHub Copilot Business and Enterprise organizations. [Request access](https://bluebear.io/contact/) · [How to enable it](#set-up-the-integration)

## Bring Copilot activity into your security workflow

Copilot is the agent most of your developers already have. Agent mode in the IDE, the Copilot CLI, and the coding agent on GitHub all run commands, edit files, and call MCP servers with the developer's access. GitHub's audit log records policy changes, seats, and the coding agent's actions on GitHub; by design it does not contain what happens in the IDE or the terminal. The policies that matter drift after they are set.

Bluebear connects in two places. Through a GitHub App it reads Copilot policies, seats, and the audit log, checks the policies against a secure baseline, and follows the coding agent's sessions and pull requests. Through the endpoint handler on the developer machine it sees what Copilot's agent actually runs and applies your policies before it does, with an approval in Slack or the terminal when a human should decide. The audit log is retrospective evidence. The handler is enforcement.

## What a finding looks like

From Bluebear's technical risk brief, *Ten ways AI agents cross runtime boundaries*.

A developer asks the agent to explain what a repository does. The agent sets up the project, runs the dependency installation, and pulls in a malicious package, even though nothing in the read-only goal required an install. Lifecycle scripts run in the developer's context, with reach to source code, SSH keys, and cloud credentials.

Bluebear connects the read-only ask to the package installation, holds it, and scopes which sessions and endpoints the component touched. Software inventory tools can identify the component; they cannot connect a read-only task to an install or say which agent session it ran in.
## What you get

- **Visibility.** Organization-wide Copilot use by seat, editor, and activity, plus every coding agent session on GitHub, in the same console as your other agents.
- **Runtime control.** Policy enforcement, approvals, and sandboxing for Copilot agent mode and the CLI on the developer machine, where GitHub's audit log does not reach.
- **Posture.** Copilot organization and enterprise policies checked against a secure baseline, with an incident for every drift and automatic resolution when it is fixed.
- **Data protection.** Secrets, PII, and source code leaving the machine through agent tool calls, held before they go.

## Set up the integration

You need a GitHub organization with Copilot Business or Enterprise, an organization owner role, and an administrator role in Bluebear.

1. In the Bluebear console, open **Settings → Integrations → GitHub Copilot** and click **Install on GitHub**.
2. Approve the read-only permissions for the organization, then click **Test Connection**.

Findings start within the hour. The integration is read-only against GitHub settings, and Bluebear keeps findings, not diffs. The endpoint handler is deployed separately from the Bluebear console.

## When you have guardrails on a bridge, you can run.

GitHub tells you which policies are set and what the coding agent did on GitHub. Bluebear tells you what Copilot's agent did on the machine, holds the actions that need a human, and alerts when the policies drift. [Request access](https://bluebear.io/contact/) and we will set your organization up.