Coding agents

A coding agent runs on a developer's laptop, not in your cluster. It reads files, runs commands and calls the network on that developer's behalf. The collector governs it there, on the machine, through the agent's own hook points.

How it works

• Decide before it runs. The agent calls a hook before every tool use. The collector answers allow or deny from your workspace policy in well under a millisecond on a warm path, and from a signed, locally cached policy bundle when the daemon is not running. No bundle and a closed fail mode means deny, with the reason shown to the developer.
• One identity per install. Every session on a machine is filed under one install identity: a hash of the machine and the OS user, never a hostname or a username in the clear. Ten repositories on one laptop are one agent in your identity page, not ten.
• Live evidence, then the transcript. The hook emits each tool call as it happens; the transcript tailer fills in what the hook could not see. The same action reaches the platform once.
• The developer's own key. A durable key bound to the install, minted through the normal login and renewed by running it again. Revoke it from the API keys page like any other key.

Supported today

• Claude Code — hooks, transcript, managed rollout for a fleet.

Other agents that expose a pre-action hook plug into the same collector inputs; see collector inputs.

What you see in the platform

• The install as one agent, with its runs, steps and the policy verdict on each tool call.
• Coverage by channel: a step seen by the runtime hook ranks above one read from a transcript.
• Drift alerts and the attestation carry how the evidence was seen, so an auditor can tell live enforcement from after-the-fact reading.