# What is collected

> The 26 device signals Holistic AI EndLayer collects, what each group reports, what the connections read, and what stays on the device at every setting.

Holistic AI EndLayer, the Guardian Agent Platform, collects metadata first, and collection is bounded. The Layer 1 Agent never elevates and never runs what it finds, so where it cannot measure something the signal says unknown instead of guessing.

## From the device: 26 signals

The agent reports 26 signals every 15 minutes, in seven groups. Each signal is listed on the [signals page](https://www.endlayer.ai/signals).

| Group | What is reported | The boundary |
|---|---|---|
| **Tools and agents** | Which of the 22 recognised AI tools, and any you add, are installed; whether each can act without asking; the Guardian plugin; what lets an agent drive a browser or desktop | Tools are never run, and tool arguments are never read |
| **Apps and web** | Which AI domains are visited per browser profile, whether your managed blocklist is in force, and the wider software inventory | No full web addresses, page titles or non-AI browsing |
| **MCP and access** | Each MCP server's name, client, transport and host; skills and commands by name; whether a provider key exists | No server is contacted, and secret values are discarded on the device |
| **Accounts and spend** | Day, tool, model and token counts, and whether an AI account is corporate or personal | No cost is calculated on the device |
| **Projects and data** | Which repositories agents work in, and data files by name only | No source code, branch names or file contents |
| **The device** | Make, model, operating system, management enrolment, updates, patch level, and the result of any approved removal | The raw hardware identifier stays on the device |
| **Security posture** | Disk encryption, screen lock, firewall, whether an EDR tool is present, network configuration, password policy and local administrators | No passwords and no traffic capture |

## Kept on the device, at every setting

Switching every option on still leaves these on the device:

- **Tool calls and their results:** a tool result can carry a whole file, so the content of tool calls is never collected.
- **Matched text:** a detection reports the detector, the action, a count and a confidence score. Where evidence is switched on, it is the prompt with every match masked on the device.
- **Browsing detail:** full web addresses, query strings, page titles and non-AI browsing history.
- **Secrets:** credential values and tokens are discarded on the device.
- **Code:** source code, branch names and where a repository sits on disk.
- **Documents:** uploaded and downloaded documents are checked on the device; their text is never sent.
- **Device secrets:** passwords, packet capture and raw hardware identifiers.

Some facts about people **are** collected, and are worth naming in an employee notice: the signed-in person's login, display name and email address as the directory knows them, local administrator account names, and install paths that can include a home folder. Identity details are visible only to administrators.

Conversation content, session titles and skill contents are separate settings, and all are off by default. See [privacy and retention](https://www.endlayer.ai/docs/privacy).

## From the connections you enable

Each connection reads **your own** account, with a credential you provide. Connections read governance metadata and usage, not prompts, responses or payloads.

| Connection | What it reads |
|---|---|
| **Microsoft Intune and Entra ID** | Managed device records including phones and tablets, device groups, rollout results, people and departments, and the apps your tenant has granted access |
| **Microsoft 365 Copilot** | Seats and use, and prompt and answer counts with the DLP detectors matched |
| **Copilot Studio and Agent 365** | Agents, their owners and access, their tool calls, and the agent registry |
| **Jamf and SCCM** | Device inventory and the groups a rollout targets |
| **Google Cloud and Vertex AI** | Model deployments and agents, graded for exposure, with the option to stop one |
| **Amazon Bedrock AgentCore** | Runtimes, gateways and their activity, through a role in your account |
| **Microsoft Foundry** | Deployments, content filters, access, usage and cost, read-only |
| **Defender, Purview, Zscaler, ServiceNow** | Security and DLP alerts, as events, without content |

Serial numbers are stored as a keyed hash, scoped to your organisation. Connection credentials are encrypted, and no part of the console or its interfaces returns them.

> **Planned:** SDKs for your own applications are planned, along with more cloud sources. The [roadmap](https://www.endlayer.ai/roadmap) lists each one.

  - [Signals](https://www.endlayer.ai/signals): All 26 signals, one by one.
  - [Privacy and retention](https://www.endlayer.ai/docs/privacy): What is on by default, and how long it is kept.

Source: https://www.endlayer.ai/docs/collection · Updated 2026-11-20
