AI agent inventory

AI agent inventory: know which agents run in your company

You cannot approve, block or audit an agent you have never listed. Most companies can name their SaaS apps but not the agents acting inside browsers, terminals and tenants today.

One inventory record example
Tool
Claude Code
Agent type
Command-line coding agent
Identifier
@anthropic-ai/claude-code (npm package)
Found by
Developer laptop package inventory
Reaches
Source code, shell, local secrets
Owner
Head of engineering
Decision
Allow on engineering devices, block elsewhere
Next review
In 90 days

Next step

AI Agent Blocklist: the reference list your inventory is matched against

Agent domains plus extension IDs, bundle IDs, Windows package IDs, process names and CLI package names, each with a category and risk grade.

From $99 a month for one organisation

Why an AI agent inventory is harder than a software inventory

Your existing asset tools were built for installed programs. Agents arrive in shapes those tools only half see, and they change names faster than any spreadsheet.

They install per profile

An agent extension lives inside one browser profile. Two profiles on one laptop can hold two different sets.

They install per user

Coding agents arrive through npm, pip or Homebrew into a user folder. Add or remove programs never sees them.

They hide in config files

An MCP server is often one block of JSON in a settings file. No installer, no entry in any app list.

They run in the cloud

Hosted agents leave nothing on the device. You only see a domain in DNS logs and an OAuth grant in identity logs.

They change names

OpenClaw was called Clawdbot and then Moltbot within weeks in early 2026. Its older macOS bundle ID bot.molt.mac still carries the old name.

They are built by staff

Agents created in your productivity suite are shared from person to person before any review.

The fix is to inventory by identifier, not by name. An extension ID, bundle ID or package name stays stable through rebrands and store listings that change their titles.

What an AI agent inventory should record

Keep the record small enough that people fill it in. These ten fields are enough to decide, enforce and audit.

FieldWhy it mattersWhere it usually comes from
Tool nameWhat people call it in tickets and requestsReference list, store listing
Agent typeBrowser agent, extension, desktop, coding, MCP, platform, tenant agentReference list category
IdentifierThe value you can block or allow onExtension ID, bundle ID, package ID, process name, domain
Identifier typeTells you which tool enforces itReference list column
Found onWhich devices, users or networksMDM, endpoint, DNS and identity logs
ReachesMail, files, CRM, code, shell, paymentsOwner interview, OAuth scopes, extension permissions
OwnerOne named person who answers for itRequester or team lead
Risk gradeSorts the review queueReference list, then your own adjustment
DecisionBlock, monitor or allow, with the reasonSecurity review
Review dateStops approvals living foreverYour policy

If you also track which accounts and tokens each agent holds, add the fields from our page on AI agent identity and permissions.

Where agents show up: the sources to pull

Each source sees part of the picture. Together they cover devices, browsers, networks and your tenant. Console menus move, so check the current admin console if a path differs.

Intune discovered apps

Apps > Monitor > Discovered apps > Export

Lists detected apps with name, platform, version, device count and publisher. On company-owned devices it collects all apps; on personal devices only managed ones.

Win32 data refreshes every 24 hours, most other types every seven days. Microsoft points to the newer App inventory for Windows as the long-term replacement.

Desktop agentsAI browsersNo extensions or CLI packages

Defender browser extensions

Endpoints > Vulnerability management > Inventories > Browser extensions

Shows extensions in Edge, Chrome and Firefox on Windows, with devices, users, versions and a permissions risk column. Advanced hunting exposes the same data in the DeviceTvmBrowserExtensions table.

Agent extensionsNeeds Defender Vulnerability Management

Google Admin console

Devices > Chrome > Reports > Apps & extensions usage

Lists extensions on managed browsers, profiles and ChromeOS devices with app ID, install type and permission count. You can block an item straight from the report's Action menu.

Agent extensionsChrome browser reporting must be on

Microsoft 365 agent registry

Microsoft 365 admin center > Agents > All agents > Registry

Lists tenant agents by type: Microsoft agents, external partner-built agents, published by your org and shared by creator. Cards show agents without owners and unmanaged agents.

Tenant agentsExport and Graph API

Mac MDM inventory

Jamf Pro, Kandji or Mosyle: installed applications report

Mac inventories record bundle IDs, which match cleanly against the macOS identifiers in an agent list. Add a check for apps that hold screen recording or accessibility rights.

Desktop agentsAI browsers

DNS, firewall and identity logs

Resolver or DNS filter logs, firewall URL logs, OAuth consent logs

The only sources that see hosted agents and personal devices on your network. Match queried domains to agent domains and review new consents with mail or file scopes.

Hosted agentsMCP servicesAny device

Personal devices are the weak spot. Intune collects only managed apps from personally owned devices, so a staff phone or home laptop shows almost nothing. For those, rely on network logs while they use your Wi-Fi or VPN, and on identity logs that show which agents hold access to work accounts.

Plan for missing data: new devices can take up to seven days to report Win32 apps when no Entra ID user has signed in, and Chrome usage data can lag by a day. Note the export date on every source.

Intune local AI agent inventory, step by step

Intune covers the installed side of the picture. These steps turn its exports into an agent list you can act on.

  1. Export discovered apps. In the Intune admin center open Apps > Monitor > Discovered apps and select Export. Keep the CSV with name, publisher, platform and device count.
  2. Match against agent apps. Join the export to the tool names and Windows package IDs in the agent list. Keep matches with their category and risk grade.
  3. Add extensions. Discovered apps does not list browser extensions. Pull them from Defender or from the Google Admin report and match on the 32-character extension ID.
  4. Add command-line agents. On developer machines, collect global npm and Python packages with a script your endpoint tool runs, and match package names such as @anthropic-ai/claude-code.
  5. Drill into devices. For each match, open Devices > All devices, pick the device, and use Discovered apps under Monitor to see what else sits next to it.
  6. Enforce. Turn block decisions into Intune policy. Our Intune guide for AI tools and extension blocking in Intune cover the policy side.

Microsoft AI agent inventory: what the tenant registry covers

What you get in the registry

  • Filters for status, publisher type, channel, platform and data source
  • Channels such as Copilot, Teams, Outlook, Microsoft 365 apps and SharePoint
  • Export to CSV with owner, publisher, platform, created and modified dates
  • Graph API access with the CopilotPackages.Read.All scope for scripted pulls
  • Block, remove or delete actions, depending on agent type

What you still need to add

  • Agent extensions and AI browsers installed by users
  • Desktop agents and personal agents on laptops
  • Coding agents and MCP servers on developer machines
  • Hosted agent platforms reached with a work account
  • Researcher and Analyst, which Microsoft treats as part of core Copilot chat rather than as agents

Watch for duplicates: Microsoft notes that one Copilot Studio agent can appear twice, as Shared by creator and as Published by your org. Treat them as one tool with two lifecycle records.

Tenant settings that limit who may install or share agents are covered on disable Copilot actions and agents.

From raw exports to a working inventory

Collect Normalise Match Grade Assign owner Decide Enforce

Normalising means lower-casing IDs, splitting extension IDs per browser and stripping version numbers from package names. Matching is then a simple join.

Example: part of an inventory for a fictional 300-person firm

ToolIdentifierTypeOwnerDecision
Claude CodeAnthropic.ClaudeCodeWindows package IDHead of engineeringAllow engineering only
Claude Code for VS Codeanthropic.claude-codeVS Code extensionHead of engineeringAllow engineering only
Claude desktopcom.anthropic.claudefordesktopmacOS bundle IDNone yetMonitor pending review
OpenClawai.openclaw.macmacOS bundle IDNoneBlock personal agent
OpenClaw (old name)bot.molt.macmacOS bundle IDNoneBlock same tool

The decisions above are an illustration, not advice about any product. Your own review sets them. For a structured review, use our AI agent risk assessment.

AI agent inventory management: keeping it current

An inventory decays the day you finish it. New agents, extensions and MCP servers appear every day, so the rhythm matters more than the first scan.

Daily

Pull the updated agent list so new domains and identifiers reach your filters and match jobs.

Weekly

Re-run the match on fresh exports and DNS logs. Triage anything new by risk grade.

Monthly

Check owners. Use the registry card for agents without owners and reassign or block.

Quarterly

Recertify every allowed agent. No answer from the owner means it moves to block.

Do not forget MCP servers

MCP servers connect assistants to your systems and are easy to miss. Keep a separate allow list of approved servers and match configs against it.

See building an MCP allowlist.

Start from what is already hiding

If you have never looked, run a discovery pass first. Our page on shadow AI agents has a 30-day plan for the first sweep.

Questions to ask every agent owner

A match tells you an agent exists. A ten-minute talk with the person who uses it tells you whether it can stay. Ask the same questions every time so answers are comparable.

What it does

  • Which task does the agent do for you, and how often?
  • Does it only read, or does it also send, buy, delete or run commands?
  • Does it run while you watch, or on its own in the background?
  • Who else uses it or relies on its output?

What it can reach

  • Which accounts did you sign in with or connect to it?
  • Can it see customer, staff, patient or student data?
  • Does it hold tokens, API keys or saved passwords?
  • Where does it send screenshots, files or page content?

Five mistakes that leave gaps in an agent inventory

Matching on names

Store titles change and copycats reuse famous names. Match on identifiers and keep names only as labels.

Only scanning Windows

Developers and executives often use Macs. Bundle IDs from your Mac MDM belong in the same match job.

Leaving owners blank

An agent without an owner has nobody to answer for it. Blank owner should mean block, not later.

Ignoring hosted agents

Cloud agents leave no files. Without DNS and consent logs they never enter the list at all.

One scan a year

Agents move faster than audit cycles. A list that is not re-matched weekly drifts out of date.

Inventory without enforcement

A list that does not feed your firewall, DNS and MDM rules documents the problem without fixing it.

How we solve it: a reference list that makes matching possible

An inventory is only as good as the list you match against. Ours names the tool behind each identifier, so a random-looking extension ID turns into a decision.

What you exportedWhat it looks likeWhat the list adds
Browser extension inventory32-character extension IDsTool name, category, risk grade for Chrome, Edge and Firefox agent extensions
Mac application inventoryBundle IDsWhich bundle IDs belong to AI browsers and desktop agents
Windows application inventoryPackage IDs and process namesMatches for agent apps and AI browsers on Windows
Developer package listsnpm, Python and Homebrew namesCommand-line coding agents and agent frameworks
DNS and firewall logsDomains2,600+ agent domains by category, from AI browsers to MCP servers

Every entry is checked before it enters the list, and the lists update every day. The CSV and JSON files carry every column, so a join in a spreadsheet or a short script is all the matching needs.

When you are ready to enforce, the same download includes firewall, DNS and policy formats; see what is in the AI Agent Blocklist.

Coding agents need their own rules on developer machines; our page on blocking Claude Code shows one worked example.

Turn your exports into an agent inventory

  • 2,600+ AI agent domains and 5,300+ device identifiers
  • Tool name, category, identifier type and risk grade on every row
  • CSV, JSON, domain lists, firewall and DNS formats, browser and MDM policy files
  • Updated every day; download access set up within one business day of your order
From $99 a month for one organisation See plans Ask a question

Ask us or order a plan

Tell us what you need to block and where your rules live. We reply within one business day.

  • Files for firewalls, DNS filters, secure web gateways, MDM and browser policies
  • Updated every day as new AI browsers, agents and extensions appear
  • Plans for one organisation, for MSPs and for product makers, or a hosted filter we run for you

We use your details only to answer you. No newsletters.

AI agent inventory: common questions

What is an AI agent inventory?

It is a list of every AI agent that runs in or for your organisation: browser agents, agent extensions, desktop agents, coding agents, MCP servers, hosted platforms and tenant agents. Each record names the tool, its identifier, owner, reach and decision.

How do I build a local AI agent inventory with Intune?

Go to Apps > Monitor > Discovered apps and export the list. On company-owned devices it includes all installed apps. Match names and publishers against agent apps, then add extensions and command-line agents from other sources, because Discovered apps does not list them.

Does Microsoft have an AI agent inventory?

For agents in your Microsoft 365 tenant, yes: Agents > All agents > Registry in the Microsoft 365 admin center, with filters, an export and a Graph API. It does not list agent extensions, AI browsers or desktop agents that users install themselves.

What should each inventory record contain?

Tool name, agent type, an enforceable identifier and its type, where it was found, the owner, what it can reach, the risk grade, the decision (block, monitor or allow) and the next review date.

How often should an AI agent inventory be updated?

Re-match at least weekly and review owners monthly. New agents, extensions and MCP servers appear every day, so a quarterly inventory is out of date almost as soon as it is finished.

Is this about AI agents that manage stock inventory?

No. Searches such as inventory aging advisor or inventory reservation assistant refer to AI features that help manage warehouse stock in business software. This page is about listing the AI agents themselves so you can govern them. For the rules side, see AI agent governance.

Can we inventory agents on personal devices?

Only in part. MDM collects little or nothing about unmanaged apps on personal devices. Network logs still show which agent services are reached from your networks, and identity logs show which agents were granted access to work accounts.

Microsoft, Microsoft 365, Intune, Defender, Copilot, Copilot Studio, Teams, Outlook, SharePoint and VS Code are trademarks of Microsoft Corporation. Google, Chrome and ChromeOS are trademarks of Google LLC.

Apple and macOS are trademarks of Apple Inc. Firefox is a trademark of the Mozilla Foundation.

Claude and Claude Code are trademarks of Anthropic, PBC. Jamf, Kandji, Mosyle and OpenClaw are trademarks of their respective owners. Names are used only to identify the products discussed. AI Agent Blocklist is not affiliated with, sponsored or endorsed by these companies.