Skip to main content
Version: 11.5.0

Endpoint Guard

The Endpoint Guard is the HYPR endpoint client running on a managed workstation. It discovers the AI tools running there and reports what it sees under a signed device identity. For the tools you choose to manage, it makes sure their traffic reaches the HYPR AgentPass™ gateway instead of going to the provider directly.

You set it up on the Desktops page of the console, which calls each enrolled workstation a desktop. Start with the page's checklist. See Getting Started with Desktops.

It gives you two capabilities, each configured on the Desktops page.

CapabilityWhat it doesConfigure it in
Agent ObservabilityEnrolls desktops and reports the AI tools they run, whether or not their traffic reaches the gatewayDesktops → Installation, and Desktops → Desktop configuration → Observation. See Enrolling Desktops and Agents Observation
Managed Desktop ToolsRoutes a tool's inference through the gateway, provisions your MCP connectors into it, and blocks its direct provider egressDesktops → Desktop configuration → Configuration. See Managed Desktop Tools

Platform Support​

The client runs on macOS and Windows, with different components on each.

PlatformComponentsAgent ObservabilityManaged Desktop Tools
macOS 13 or laterA per-user background service, plus a system network extension. The extension attributes connections to the application that made them, and enforces egress blocksYesYes
Windows 11A Windows service. It also enforces egress blocksYesYes, except Claude and Codex channels

On macOS, the client ships as a signed installer package or disk image, or with HYPR Passwordless for Mac. On Windows, it ships as a single installer that a user double-clicks or a deployment tool runs silently. You download the installers from Desktops → Installation. The client needs two values at deployment time: your gateway address, and an enrollment token from the same tab. Confirm packaging for your device-management tooling with your HYPR representative. Platform Support has per-capability detail.

How It Discovers an Agent​

The client works from the observation manifest your organization publishes. The manifest lists tools, each described by the signals that identify it. The client recognizes a running program when any one of these signals matches:

SignalWhat it identifies
Process and executable namesThe running program, such as claude or codex.exe
PathsWhere the program was launched from
Command lineText in the command line. This is how the client recognizes an agent inside a general-purpose runtime such as Node.js or Python
Code signatureThe code-signing identifier on macOS, or the signed publisher on Windows. High confidence, because a look-alike process cannot forge it
Application identityA macOS bundle identifier, or a Windows packaged-app identity
Network connectionsOn macOS, the signed application that opened a connection to an AI provider's hosts
Tool eventsFor Cursor, the tool's own event hooks, which report each action

The client matches the strongest signals first, and each report records which signal matched. A program that matches no tool is discarded on the workstation. It is never recorded, queued or sent.

Adding a tool to the manifest, on Desktops → Desktop configuration → Observation, is a console change, not a client update. Workstations pick it up on their next check-in.

Discovery runs continuously while a workstation is enrolled and Watch desktops is on:

  • macOS scans the signed-in user's processes every few seconds, and reports connections and Cursor hook events as they happen.
  • Windows subscribes to process start and stop events.

On busy fleets, the manifest's sampling setting reports a deterministic fraction of matched activity.

What a Report Carries​

Each report the client sends carries these attributes.

AttributeSource
ToolThe manifest tool that matched, with its display name and category
OwnerThe user signed in to the workstation
DeviceThe enrolled workstation, established cryptographically at enrollment
ActivityStart, stop or action, with a timestamp and process identifiers
Matching signalWhich signal recognized the tool, as a measure of confidence
Command-line detailsArguments after redaction, and the working directory where available
DestinationFor a network connection, the provider host reached

Reports appear on Activity → Desktop activity. See Desktop Activity.

Observation never blocks a tool. A managed tool that has not been onboarded is still routed, and the connector's onboarding mode decides what happens when its traffic arrives.

How It Routes and Enforces​

The gateway makes the routing decision, and the workstation applies it.

You set intent in Agent Configuration: which tools are managed, which capabilities each gets, which connector each routes through, which have direct egress blocked, and how a tool behaves in an outage. The gateway resolves this into a configuration for each workstation, from platform defaults, your settings and that desktop's own settings.

The client adds one check the gateway cannot make: whether the tool is installed and signed in. It defers configuring a tool that is not signed in, and reports the deferral.

The client applies three controls:

  • Routing configures the tool itself to send inference to your gateway. It uses an environment variable, the tool's configuration file or, for Claude Desktop on macOS, managed preferences. It prefers an enterprise-managed location. Nothing is proxied. See Routing and Observation.
  • MCP provisioning adds your published MCP connectors to the tool's configuration. See Provisioning MCP Tools.
  • Egress enforcement blocks the tool's direct connections to its provider hosts. On macOS, the block is scoped to the tool's code-signing identity and the listed hosts. Sign-in endpoints stay reachable. On Windows, the block applies to every process on the workstation, browsers included. See Egress Enforcement.

Routing is applied before enforcement, so a tool is never left without a path to its provider.

The client checks in every 60 seconds by default. Each check-in fetches the manifest and, while Configure agents from AgentPass is on, the endpoint configuration. Before changing a tool's settings, the client backs them up, and it merges changes instead of overwriting.

Turning off Configure agents from AgentPass, or turning configuration off for one desktop, removes the egress blocks and what AgentPass added to each tool, on the next check-in. Each desktop's Managed configuration shows what was applied to each tool. See Managed Desktops.

Identity and Transport​

The client generates its signing key on the workstation, in the Secure Enclave or TPM where present. The private key never leaves the device. After enrollment, every call carries a short-lived signed assertion, and every activity event is a signed security event token.

The gateway verifies each signature and rejects stale or replayed events. It then reapplies your redaction rules and rate-limits each device. Rejected reports appear in the ingest-activity log.

All communication is outbound HTTPS to your gateway, and the client opens no listening ports. Revoking a desktop on Desktops → Installation rejects its next report until it enrolls again. Updating Endpoint Guard keeps the workstation's identity. Removing it retires that identity.

Agent Observability uses the OpenID Shared Signals Framework (SSF). The gateway is the receiver, and each client is a transmitter of signed events. To build your own client, ask your HYPR representative for the endpoint-client integration guide.

What Is Not Collected​

The client limits what it collects:

  • No prompts, model responses, file contents or network payloads. The network extension sees only encrypted connections.
  • Only tools in your manifest and, on macOS, only the signed-in user's processes.
  • Command-line arguments are redacted on the workstation under your redaction rules. The working directory is truncated to the depth you allow.
  • Cursor hook events are reduced to the event name, working directory, conversation identifier and redacted command.
  • The enrollment token, signing key and report contents are never written to the client's logs.