Skip to content

GUIDE, read 12 Nov 2026

Agentless on an endpoint describes how a tool is delivered, and each delivery method sees a different slice of AI activity

See the full indexAll guides

Surfaces this guide covers

  • Desktop apps
  • IDE assistants
  • Command-line agents
  • MCP servers
  • Browser extensions
  • SaaS copilots

Actions it covers

  • File read
  • Shell command
  • Network call
  • Credential access
  • Data paste
  • Package install

Agentless is a statement about install method, not about what a tool can see

When a vendor says a product is agentless, the claim usually means one thing: no new endpoint sensor ships to the laptop. It does not say where the product gets its data. A tool that reads configuration files through your MDM, a tool that lives inside the browser, a tool that calls a SaaS vendor's admin API and a tool that reads the telemetry your EDR already collects are all described as agentless. They share a label and almost nothing else.

For AI endpoint security the difference matters more than it did for classic endpoint tools. The activity you care about is spread across at least six surfaces: desktop AI apps, IDE assistants, command-line agents, MCP servers, browser extensions and SaaS copilots. No single delivery method reaches all six. A browser extension cannot see a terminal agent. An API connector cannot see a process on a laptop. EDR telemetry can see the process but usually not the prompt or the tool call inside it.

This guide lists the five delivery methods that get called agentless, says what each can and cannot see, and ends with a test you can run against a vendor in an afternoon. Vendor product claims are written as the vendor's. This page makes no claim about any specific product's coverage.

MDM-pushed profiles configure the agents that already run, but they do not watch them

A mobile device management (MDM) profile is a settings payload the operating system applies. Microsoft Intune, for example, lets you add custom settings to macOS devices by supplying an .xml or .mobileconfig file, per Microsoft's Intune documentation (read 12 Nov 2026). Windows uses registry policy or Group Policy for the same job.

This is a real control channel for AI tools. Anthropic documents that Claude Code accepts endpoint-managed settings deployed through native OS policies (macOS managed preferences, Windows registry) or a managed settings file, and that these sit above user settings in precedence (Anthropic, Claude Code server-managed settings page, read 12 Nov 2026). Claude Code also reads a managed-mcp.json from a fixed path per platform: /Library/Application Support/ClaudeCode on macOS, /etc/claude-code on Linux and WSL, and C:\Program Files\ClaudeCode on Windows (Anthropic, Claude Code managed MCP page, read 12 Nov 2026). VS Code publishes enterprise policies for MCP access and an MCP server allowlist (Microsoft, VS Code enterprise policies page, read 12 Nov 2026). Chrome and Edge take extension allow and block lists through policy.

What a profile cannot do is tell you what happened. It sets a rule and the application enforces it, or does not. If a user runs an AI tool that reads no managed policy, the profile has nothing to say about it. A profile is a good way to turn off what you have decided to turn off, and a poor way to learn what you have not yet found.

  • Sees: nothing by itself. It is a configuration channel, not a sensor.
  • Can enforce: allowlists and denylists for tools that read managed settings, extension install rules, MCP server restrictions.
  • Misses: any AI tool that ignores managed settings, any tool you did not know to write a policy for, and any action taken inside an allowed tool.

Browser extensions see the page and the clipboard, and stop at the browser edge

An extension deployed to Chrome or Edge by policy runs inside the browser with the permissions its manifest requests. Google documents that permission warnings describe the capabilities an API grants, and that a host permission such as all URLs lets an extension act on every site (Chrome for Developers, permission warnings page, read 12 Nov 2026). That is why a security extension can see prompts typed into a web chat, files chosen in an upload dialog and text pasted into a page.

Microsoft's own endpoint DLP shows the pattern. After you install the Microsoft Purview extension for Chrome on Windows 10 or 11, the organization can monitor attempts to access or upload a sensitive item to a cloud service in Chrome and enforce protective actions (Microsoft Learn, Purview extension for Chrome page, read 12 Nov 2026).

The edge of this approach is the browser itself. A coding agent in a terminal, a desktop AI app built on its own runtime, and a local MCP server make network calls and file reads without touching a page. Another browser, an Incognito window where extensions are not allowed, or a browser profile your policy does not cover will also fall outside it. Google's Chrome Enterprise help describes Incognito as a way users can escape extension-based controls unless you manage it with the IncognitoModeAvailability policy (Google, Chrome Enterprise Help, read 12 Nov 2026).

  • Sees: prompts, pastes, uploads and page content in the managed browser.
  • Can enforce: block, warn or redact at the moment of paste or upload, in the browser.
  • Misses: terminal agents, desktop AI apps, local MCP servers, other browsers, unmanaged profiles.

OS APIs give process and file events, but they need code running on the machine

Operating systems expose security events to approved software. On macOS that route is Apple's Endpoint Security framework, documented on Apple's developer site (read 12 Nov 2026). It is how a security product learns that a process started, a file was opened or an executable ran. Windows has equivalent event sources.

These APIs are the strongest source for the questions an AI endpoint team asks first: which AI binaries run on which machines, which child processes did a coding agent start, which files did it open. They do not carry the meaning of the action. A shell command started by a coding agent looks, at the process level, like a shell command started by a developer. Telling them apart means linking the process to its parent, to the application that launched it, and ideally to the agent session.

A reader of vendor material should notice one point. Anything that consumes these APIs is code on the machine, whatever the marketing page calls it. That is not a flaw. It is the price of seeing process and file events. If a vendor says agentless and also says it sees every command an agent runs, ask which component on the laptop produces that data.

  • Sees: process start, file access, network connections, depending on the sensor.
  • Can enforce: block a process or file operation, if the sensor runs in an enforcing mode.
  • Misses: the prompt and the agent's reason for acting, unless the product also reads the agent's own logs or hooks.

Existing EDR telemetry is reuse, and its depth depends on what your EDR already collects

Some vendors position themselves as agentless because they read data from the EDR you already pay for. That is a legitimate design, and it keeps the deployment small. The limit is that the product can only see what the EDR was configured to record. If your EDR keeps process starts for 30 days but does not record command-line arguments for some process classes, the AI tool inherits that gap.

Ask what the integration pulls, how often, and what it does with it. A read-only feed supports inventory and detection after the fact. It supports blocking only if the product can also send a response action back to the EDR, and that response then depends on the EDR's own latency.

Whether a given EDR is the right base for AI coverage is a separate question. Here the point is narrower: reuse of EDR telemetry is a delivery method with its own boundary, and the boundary is the EDR's data model.

  • Sees: processes, command lines, network connections and file events your EDR already records.
  • Can enforce: only what the EDR's response actions allow.
  • Misses: anything the EDR does not record, and the content of prompts and tool calls.

API connectors reach SaaS copilots and OAuth grants, not the laptop

An API connector authenticates to a SaaS service with an admin credential and reads what the service exposes. For AI, the useful data is the list of third-party apps that users have connected to a tenant. Microsoft documents how to review permissions granted to applications in a Microsoft Entra tenant, covering applications added through user or admin consent (Microsoft Learn, manage application permissions page, read 12 Nov 2026). Similar admin views exist in other identity and productivity platforms.

This is the cleanest agentless method, and it is entirely accurate to call it agentless. It shows which AI apps hold OAuth access to mail, files and calendars, and which scopes they hold. It does not show a single thing about a laptop. A local MCP server, a command-line agent that uses a personal API key, and a desktop app that never touched your identity provider all stay invisible to it.

Treat an API connector as a source for the SaaS copilot surface and for OAuth grants, and judge any broader claim against that boundary.

  • Sees: connected apps, scopes, consent events, SaaS copilot usage where the vendor exposes it.
  • Can enforce: revoke a grant, disable an app, change a tenant setting.
  • Misses: processes, files, local configs and traffic on the device.

A side-by-side view of what each method sees

The table summarizes the five methods against the six surfaces this site maps. 'Yes' means the method can observe the surface in principle. It does not say a given product does.

Delivery methodDesktop AI appsIDE assistantsCommand-line agentsMCP serversBrowser extensions and AI browsersSaaS copilots
MDM profileConfigure onlyConfigure onlyConfigure onlyConfigure onlyConfigure onlyNo
Browser extensionWeb versions onlyNoNoNoYesWeb access only
OS API sensorProcess and file eventsProcess and file eventsProcess and file eventsProcess and file eventsProcess eventsNo
EDR telemetry reuseProcess eventsProcess eventsProcess eventsProcess eventsProcess eventsNo
API connectorNoNoNoNoNoYes

A four-step test for any agentless claim

Ask the vendor for a trial on a small set of devices, and run the same four steps against every candidate. The steps use ordinary tools and need no attack content.

Step one: name the data source. For each surface the vendor says it covers, get the exact component that produces the data: an MDM payload, an extension ID, a system extension, an EDR API, a SaaS API. If the answer is a laptop component, the product has an endpoint footprint, and you should plan for it. If the answer is an API, the product cannot see the laptop.

Step two: run one tool per surface. Start a desktop AI app, a coding assistant in an IDE, a command-line agent, a local MCP server defined in a config file, an AI browser extension, and a SaaS copilot sign-in. Check which of these appear in the product's inventory within a day.

Step three: make each tool do something observable and harmless. Have the coding agent read a file in a test folder, run a benign shell command and make one outbound request. Have the desktop app upload a test file. Compare what the product recorded with what you did, including the tool call name, the file path and the user.

Step four: ask what happens when the product is turned off or removed. For an extension, can a user disable it? For a profile, can a user remove it? For a sensor, is there tamper protection? A control that the person under control can remove is a log, not a control.

Record the results in a table like the one above and score what you saw, not what the data sheet says. The scoring approach used on this site follows the same rule: a vendor's own claim is capped at 7 of 10 on a criterion until a second source confirms it, and a feature with no evidence found scores 0.

Where this sits in the index

Sources

  1. Microsoft Learn: Intune custom settings for macOS, read 12 Nov 2026
  2. Anthropic: Claude Code server-managed settings, read 12 Nov 2026
  3. Anthropic: Claude Code managed MCP, read 12 Nov 2026
  4. Microsoft: VS Code enterprise policies, read 12 Nov 2026
  5. Google: Chrome extension permission warnings, read 12 Nov 2026
  6. Microsoft Learn: Purview extension for Chrome, read 12 Nov 2026
  7. Google: Chrome Enterprise Help, extensions in Incognito, read 12 Nov 2026
  8. Apple: Endpoint Security framework, read 12 Nov 2026
  9. Microsoft Learn: manage application permissions, read 12 Nov 2026

Last reviewed 12 Nov 2026.