Learn
The EndLayer glossary.
Every term you will meet across the Guardian Agent Platform, in plain words. Each one links to the others, so you can follow an idea from a device to a decision.
⌘K
Start with the flow
How EndLayer works
Seven steps, from the agent on a device to the evidence an auditor reads. Open any step for its term.
- 01DeployLayer 1 Agent
- 02SignalsSignal
- 03InventoryAI inventory
- 04ContextContext
- 05PolicyPolicy
- 06DecisionVerdict
- 07EvidenceEvidence
Guardian Agent PlatformEndLayer's category: guardian agents that find the AI your company runs, understand it and hold it to your rules.Shadow AIAI tools and agents people use at work without the organisation knowing about or approving them.SignalOne specific thing the agent measures on a device, such as installed AI tools, MCP servers or disk encryption.Policy versionEach published edit to the policy creates a new numbered version that can never be changed.
The platform
What EndLayer is and the parts it is made of.
- Guardian Agent PlatformEndLayer's category: guardian agents that find the AI your company runs, understand it and hold it to your rules.
- Guardian agentSoftware that sits where AI is used, watches it and acts on your rules, instead of waiting at the network edge.
- Layer 1 AgentThe one small EndLayer agent installed on each device. One install does all three jobs against the same policy.
- Device agentThe lightweight app that runs quietly on macOS, Windows and Linux and reports what AI is on the machine.
- ConsoleEndLayer's web dashboard: devices, AI tools, agents, people, cost, policies and evidence in one place.
- SurfaceThe first capability: find what AI exists on your devices. Know what exists.
- MonitorThe second capability: understand how AI is used, by whom, with what reach and at what cost. Know what changes.
- GovernThe third capability: apply your rules, act where AI runs, and keep the evidence. Prove the response.
Discovery
What EndLayer finds on a device, and how.
- SignalOne specific thing the agent measures on a device, such as installed AI tools, MCP servers or disk encryption.
- ScanOne pass of the agent over a device, about every 15 minutes, that reports the current signals.
- AI inventoryA complete, current list of the AI tools, agents, models, extensions and connections on your devices.
- AI toolAn AI desktop app, command-line tool or editor extension found on a device, such as ChatGPT, Claude, Copilot or Cursor.
- Exposed keyAn AI provider key left in a settings file on a device. EndLayer reports that it exists; the key itself is discarded.
- Personal accountSomeone using a personal ChatGPT or Claude login for work instead of the company's managed account.
- AI web usageWhich AI websites are used on each browser profile. Only the AI site names are recorded.
- Device postureThe basic security health of the machine running the AI: disk encryption, screen lock, firewall, updates and protection.
- ProjectA code repository or work folder where AI agents are used. EndLayer records the project, not its contents.
Context and spend
The facts that turn a discovery into a decision.
- ContextThe facts that turn a discovery into a decision: which device, person, account, project and connections sit behind an AI tool.
- ReachHow far an AI tool extends: how many machines carry it, and what systems it can get to through its connections.
- AutonomyHow freely an AI agent can act without a person approving each step.
- OwnerThe person accountable for a flagged AI tool or finding, tracked until it is closed.
- AI spendWhat AI costs, attributed by app, agent, person and project, so spend is understood before the invoice arrives.
- Token costAn estimate of AI spend, calculated from the usage counts each device reports.
- SeatA paid licence for an AI tool, such as a Copilot or ChatGPT seat, compared with real use.
Policy and control
How your rules are written, applied and enforced.
- PolicyYour organisation's AI rules, built in the console from a library of ready-made rules and frameworks.
- Policy versionEach published edit to the policy creates a new numbered version that can never be changed.
- Re-gradingWhen the policy changes, the fleet is checked again against the new version.
- Rule libraryThe catalogue of ready-made policy rules, such as AI tool access, blocked MCP servers and sensitive data in prompts.
- SanctionedYour decision on an AI tool: sanctioned tools are approved; warn flags a tool; block also queues it for removal.
- UndecidedAn AI tool nobody has made a decision on yet.
- GuardrailAn action taken on the device as AI is used: record, warn, redact or block.
- Sensitive data in promptsAn on-device check that catches personal data, card numbers or secret keys in a prompt before it reaches an AI provider.
- DetectorA built-in check for one kind of sensitive data, such as a card number, IBAN or health number.
- Confidence thresholdHow sure a detector must be before it blocks. Below it, a match only warns.
- Kill switchAn emergency stop that halts one AI agent, app or MCP server across the fleet, or every agent at once.
- Removal with approvalThe agent removes a blocked AI app or MCP server from a device, only after an administrator approves it.
- ExceptionA dated, recorded carve-out that excuses a device or person from one rule until a set day.
- Device groupA set of devices that can be held to stricter rules than the baseline every device follows.
- InboxThe console's work queue: devices waiting to join, removals to approve and devices that need attention.
Evidence and audit
The record that shows what was found and decided.
- FindingSomething EndLayer surfaced that needs a decision, such as an unknown agent, a risky MCP connection or an exposed key.
- VerdictThe result of checking one item on one device against your policy: pass, warn or fail.
- UnknownThe device could not measure something. It is shown separately and never counted as a pass.
- UngradedThe fact was measured, but your policy does not say anything about it yet.
- ControlA governance requirement, phrased the way auditors talk, that one or more signals provide evidence for.
- FrameworkA recognised standard your policy is mapped against, such as SOC 2, the EU AI Act, ISO/IEC 42001, NIST AI RMF or GDPR.
- CoverageHow much of a framework or a fleet your policy and devices actually address.
- EvidenceThe detail behind each result: what was found, on which machine, when, and under which policy version.
- Evidence packA ready-made export of evidence for a chosen period, designed to be handed to an auditor.
- Audit trailA permanent record of every policy version and every administrative action, with who acted, when and why.
- EventA record of something that happened on a device, such as a prompt that was warned on, redacted or blocked.
- ReportA one-click document for a particular reader, such as an executive summary, a CISO review or a cost review.
Deployment and privacy
How EndLayer is rolled out, and what it never collects.
- EnrollmentHow a device joins your organisation in EndLayer.
- MDMMobile device management: the tools IT already uses to install software, such as Intune, Jamf or SCCM.
- UnprivilegedThe agent runs as the signed-in user, never asks for admin rights and never runs what it finds.
- Metadata firstBy default EndLayer collects facts about AI use, not what people typed or what the AI replied.
- RedactionRemoving sensitive values, such as a key or a card number, so they never leave the device or get stored.
- PauseAn administrator can stop a device scanning for a set time, with a reason shown to the user.
- HostingWhere EndLayer runs: in its hosted console, or on your own infrastructure by arrangement.
AI concepts
General AI terms you will meet across the platform.
- Shadow AIAI tools and agents people use at work without the organisation knowing about or approving them.
- AI agentAI software that takes actions on its own, such as editing files, running commands or calling other systems.
- Coding agentAn AI agent that writes and changes code on a developer's machine, such as Claude Code, Cursor, Codex CLI or Copilot.
- MCPThe Model Context Protocol: an open standard that lets AI tools plug into other systems, such as code hosts, databases and apps.
- MCP serverOne connection given to an AI tool through MCP, such as access to a code host or a database.
- Local modelAn AI model that runs directly on a laptop, for example through Ollama or LM Studio.
- SkillA packaged set of instructions an AI agent can run on request, often shipped inside a code repository.
- Permission modeA setting that decides whether an agent must ask before each action, or may act alone.
- Prompt injectionHidden instructions inside a web page, document or tool result that trick an AI agent into doing something its user never asked for.
- TokenThe unit AI providers measure and bill usage in, roughly a piece of a word.
- Network gatewayA traditional control that inspects traffic sent through a central checkpoint. EndLayer works on the device instead.
EndLayer
Give AI room to work.And clear boundaries.
See the Guardian Agent Platform on your own fleet. Find every AI tool, understand what it reaches and costs, and apply the controls your team needs.
One connected layer of control
Devices · Cloud · ApplicationsSurface / Monitor / Govern