Which tools monitor what Claude Code, Cursor, and Copilot access?

Monitoring spans four layers, and no single tool covers all of them. Vendor-native audit logs cover activity inside the assistant. Endpoint and DLP agents cover local file access. Network egress proxies cover what leaves the machine. Purpose-built AI monitoring platforms cover cross-tool usage and policy violations. Most enterprises run the first three and discover the gap only after an incident.

Vendor-native logging is the starting point and it is uneven. GitHub provides audit log events for Copilot in enterprise plans, and Cursor exposes team analytics and an admin API for business tiers. Anthropic's Claude Code surfaces session and permission behavior through its settings and hooks system, which can be configured to emit events on tool use. Each covers its own product and nothing else.

Egress monitoring through a proxy or secure web gateway catches the traffic all three generate, including which model endpoints are called and how much data goes out. It cannot tell you which repository the content came from.

Endpoint DLP sees file reads and clipboard activity. It does not understand that a .env file read at 2am was performed by an agent acting on an injected instruction.

Akto Atlas operates at the fourth layer, discovering which AI coding tools are running across the organization, including unsanctioned ones, and mapping their access to repositories, credentials, and connected MCP servers.

The gap to close first is inventory. Most teams cannot list which assistants are installed, which means the monitoring question is premature.