# Trust model

> What the Holistic AI EndLayer agent can do on a device, the promises built into its design, how devices prove who they are, and how releases are signed.

Holistic AI EndLayer, the Guardian Agent Platform, puts an agent on people's devices, so a security reviewer's first question is what that agent can do. This page answers it. For Holistic AI's certifications, policies and subprocessors, see the [Trust page](https://www.endlayer.ai/trust).

## Seven promises built into the design

Each is a property of how the Layer 1 Agent is built, not a setting somebody can change, so a reviewer has no configuration to audit for them.

| Promise | What it means |
|---|---|
| **It never elevates** | The agent runs as the person whose device it is. Where that means it cannot read something, the signal says unknown rather than guessing. |
| **It never runs the tools it finds** | Finding Claude Code does not involve running it. Even a version number is read from the installation. |
| **It never connects to what it finds** | MCP servers are read from their declarations. None is contacted. |
| **Nothing leaves until it is pointed at your organisation** | Until the agent is connected to your EndLayer service, it sends nothing. |
| **Certificate checks stay on** | Certificate verification is always on. Your own root certificate can be added for TLS inspection. |
| **Its credential is kept out of its configuration** | The device's own credential is stored separately from the settings file and is never passed on a command line. |
| **Conversation content is off by default** | Reading what people typed takes a deliberate decision, a stated lawful basis, and an audit record on every read. |

## What the agent changes on a device

Collection changes nothing. The agent changes a device only when a published policy, or an administrator's decision in the console, asks it to:

- **Prompt scanner:** it adds its scanner hook to Claude Code, Cursor and Codex CLI, and leaves the person's own hooks as they were.
- **Browser extension:** it registers itself so the extension on the same device can reach it.
- **Guardian plugin:** it places the plugin in the agent hosts your policy names.
- **Removal:** it removes an app or MCP server an administrator approved for removal, using the person's own permissions.

Hooks and plugins are taken out again when the policy stops asking for them. A removal stays removed.

## How a device proves who it is

Each enrolled device ends up holding a credential of its own. A shared secret is never a device's identity.

- **One device:** the agent shows a short claim code after `endlayer enroll`, and somebody signed in to your organisation approves it in the console. The person approving decides which organisation the device joins.
- **A fleet:** one enrolment secret, with a device limit and an expiry you set, delivered in a file by your device management tool rather than on a command line.

Each credential is scoped to one job:

- **A fleet enrolment secret:** enrols devices, within the limit and expiry you set.
- **A device's credential:** sends that device's own reports.
- **Joining an organisation:** always needs a person in your organisation to approve it.

Revoking a fleet secret stops new enrolments and leaves enrolled devices reporting, so rotating it causes no outage.

## Tenant isolation

Each organisation is its own tenant. The organisation a request acts on is always worked out on the server from the signed-in caller, never taken from the request itself, so no request can ask for another organisation's data.

## Who can do what

Members see the fleet; administrators also see who, and manage settings; owners also manage billing. See [privacy and retention](https://www.endlayer.ai/docs/privacy). Every administrative action, including publishing policy, approving a removal, granting an exception and tripping a kill switch, is in the audit trail with who, when and the reason given.

## Supply chain

- **Dependencies:** pinned and licence-checked on every build.
- **Releases:** each one ships a signed checksum manifest and build-provenance attestations.
- **Windows:** installers are Authenticode signed.
- **macOS:** installers are signed with Holistic AI's Developer ID and notarised by Apple.
- **Downloads:** installers come from your own organisation's console, and every download can be checked against the signed manifest.

Build provenance is available on request. For certifications, reports and policies, see the [Trust page](https://www.endlayer.ai/trust).

## Planned

- **Approval before an agent acts:** see [what it stops](https://www.endlayer.ai/docs/enforcement).
- **SDKs for your own services:** see the [roadmap](https://www.endlayer.ai/roadmap).

Self-hosting is available by arrangement. See [where it runs](https://www.endlayer.ai/docs/hosting).

  - [What is collected](https://www.endlayer.ai/docs/collection): Each signal group, and what is never read.
  - [Privacy and retention](https://www.endlayer.ai/docs/privacy): Content settings, retention and roles.

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