Skip to main content

The Guard Cascade

Guard is the policy every managed device enforces — what the browser extension captures and blocks, what coding-agent hooks allow, ask about or deny, and what is contained on the machine. You manage it under Endpoints → Guard.

There is one Guard, built from layers. Each layer sets only the fields it cares about; more specific layers override less specific ones field by field. Set a sensible baseline once for the workspace, then refine it for a kind of app, a team or a single device without restating everything.

The cascade​

Global  →  App type  →  Team  →  Device  →  Agent group  →  Agent
(least specific) (most specific)

For each field, the value from the right-most layer that sets it wins; anything not set falls through to the layer on its left, and finally to Waxell's built-in defaults.

LayerApplies to
GlobalEvery device in the workspace — your baseline.
App typeOne kind of AI app wherever it runs, e.g. Claude Code or Cursor.
TeamDevices whose owner belongs to a team.
DeviceOne machine.
Agent groupA group of governed agents.
AgentOne governed agent.

List fields (blocked commands, extra hosts…) are replaced, not merged, by a more specific layer. Hook settings are the exception: they merge per tool.

The built-in defaults​

With no layers at all, a device runs Waxell's recommended defaults:

  • Browser capture on, secrets redacted on the device; browser DLP enforcement off.
  • Coding-agent hooks on for every supported tool, with the recommended blocked commands, protected files and git protections.
  • Containment off — nothing is blocked on the device until you add it.

The What this fleet runs panel on the Guard page shows the effective baseline at any time.

The Guard editor​

Open any layer to edit it. The editor has these tabs (each shows how many fields this layer overrides):

TabWhat it controls
Guard RulesWhat coding agents may do: commands, files, git, agent DLP, session limits, infrastructure files, endpoint lock-down. Built from a catalog of named rules you can search and tick.
Browser CaptureBrowser capture upload, extra AI sites, browser DLP, file-upload scanning. See Browser Capture.
ChatGPT DesktopAfter-the-fact capture of the ChatGPT desktop app's chat.
NetworkWhich domains and internal addresses coding agents may fetch.
ContainmentSite, vendor, app and user blocks, and quarantine. See Containment.
HooksWhich coding tools are hooked, events, matchers and time budgets. See Coding Agents (Hooks).
EffectiveThe fully resolved policy, with the layer each field came from.
SourceThis layer's own settings as JSON.

Templates. A new, empty layer opens with a gallery of starting points — fill the layer from one and adjust.

Check a policy before you rely on it​

Resolve a context on the Guard page takes an app type, team, device and agent and shows the effective policy with a badge on every field naming the layer it came from. Use it to answer "if someone in Engineering runs Cursor on this laptop, what happens?" without merging layers in your head.

How policy reaches devices​

Each device checks for a changed policy about once a minute and applies it immediately — no MDM re-push. Enforcement runs from the copy on the device, so it keeps working offline. One exception: adding Browser extra hosts on macOS also needs the browser profile re-pushed, because the browser itself must allow the new site.

Export and import​

Export downloads your Guard — global, app-type and team layers, agent and agent-group layers — as one file. Import loads it into another workspace: it shows a plan first, checks every field, creates missing agent groups, reports agents it cannot find, and never copies device layers. Importing the same file twice changes nothing. Useful for MSPs rolling one standard out to every client.

Next​