The agent

The agent is one process per machine. It captures AI interactions using the strategies its workspace enrolled it with, and reports what it captured to the console.

The five strategies, at a glance

A machine enrols with a set of these — see Capture strategies for how each one captures, what it cannot see, and its named limits.

• TLS interception — decrypts traffic at a local listener, using a certificate made for this machine.
• Network steering — sends browser and proxy-aware traffic to that listener, with a proxy setting and a claimed route.
• CLI & IDE tools — asks each tool's own hook before it acts, and answers it.
• Browser extension — reads prompts inside the browser page itself.
• File inspection — scans configured folders on disk for AI-related files.

What it never does

• A machine enrolled with only CLI & IDE tools and/or the browser extension binds no port and claims no network route — nothing for a customer's VPN, SASE agent, or EDR to conflict with.
• Network steering alone is never given a certificate authority. Only TLS interception mints one, and only when it is enrolled too.
• Claude Code cloud sessions never touch the machine's network. No strategy here can intercept them.
• Under CLI & IDE tools, a hook is asked before a tool acts and nothing is re-routed: a developer's own API key never leaves the tool.