---
title: Which AI asset types should an AI inventory cover?
url: https://opencontroller.lyzr.com/blog/ai-asset-types-to-inventory
published: 2026-11-12
updated: 2026-11-12
author: Lyzr
---

# Which AI asset types should an AI inventory cover?

**An AI inventory that lists only agents misses most of the risk. Cover ten asset types: agents, models and endpoints, API keys, tools and MCP servers, skills, workspace assistants such as custom GPTs, coding agents, prompts and system instructions, vector stores, and the identities agents run as.** Each one lives in a different place, so each needs its own discovery method.

This is the question buyers ask right after "can you find our agents?": what else gets found? This post answers it as a checklist, with where each asset type usually lives. For the full discovery method, see the pillar guide on [how to find every AI agent in your enterprise](/blog/how-to-find-shadow-ai-agents).

## Why is "agents only" not enough?

An agent is a composition. It is a model, called with a credential, given instructions, wired to tools, sometimes grounded on a vector store, and running as some identity. Risk can enter through any of those parts.

- A leaked API key lets anyone spend on your account, agent or not.
- An unreviewed MCP server can give a well-behaved agent a destructive tool.
- A skill uploaded by one employee can change how hundreds of assistants behave.
- A vector store can hold regulated data that the agent's own permissions would never allow.

NIST's AI Risk Management Framework makes inventory a governance outcome in its own right: GOVERN 1.6 calls for mechanisms to inventory AI systems, resourced according to organisational risk priorities. In practice, "AI systems" means the parts above, not just the top-level agent.

## What are the ten AI asset types to inventory?

| # | Asset type | Where it usually lives | Key fields to record |
| --- | --- | --- | --- |
| 1 | Agents | Cloud agent services (Amazon Bedrock and Bedrock AgentCore, Microsoft Foundry, Vertex AI), SaaS platforms (Salesforce, SAP, ServiceNow, HubSpot), custom code on Kubernetes | Owner, platform, version, status, business purpose |
| 2 | Models and endpoints | Cloud model catalogues, self-hosted inference servers, data platforms | Provider, model ID, region, who can invoke |
| 3 | API keys and credentials | Provider consoles, secrets managers, CI variables, developer laptops | Issuer, scope, last used, rotation date, owner |
| 4 | Tools and MCP servers | Agent configs, internal registries, local config files such as `.vscode/mcp.json` | Endpoint, auth method, tools exposed, write capability |
| 5 | Skills | Claude Team and Enterprise workspaces, agent frameworks, repos | Author, scope (org or personal), scripts included |
| 6 | Workspace assistants | ChatGPT Enterprise custom GPTs, Claude Projects, Copilot Studio agents | Creator, sharing level, connected data |
| 7 | Coding agents | Developer devices and IDEs (Claude Code, Cursor, GitHub Copilot) | User, device, MCP servers, permission mode |
| 8 | Prompts and system instructions | Code repos, prompt management tools, platform configs | Version, linked agent, last change |
| 9 | Vector stores and knowledge bases | Managed vector databases, search services, Postgres with pgvector | Data classification, source systems, who can query |
| 10 | Agent identities | Identity provider, cloud IAM roles, service accounts | Permissions, on-behalf-of scope, human owner |

## How should you prioritise when you cannot do all ten at once?

Start where spend and blast radius are highest. A reasonable order:

1. **Keys and models.** They are the easiest to find (provider and cloud APIs) and they show where money goes.
2. **Agents and their identities.** Link every agent to a named human owner and to the role it runs as.
3. **Tools and MCP servers.** Record which ones can write, delete, send or pay. Those need review first.
4. **Coding agents and local MCP servers.** These sit on laptops, outside cloud discovery, so they need device-level coverage.
5. **Skills and workspace assistants.** Pull them from the SaaS workspace admin APIs where available.
6. **Vector stores, prompts.** Map them once the agents that use them are known.

## What makes an inventory entry useful rather than a list?

An entry is useful when it answers four questions without a meeting: who owns it, what it can touch, what it costs, and how to stop it. That means each record needs relationships, not just a name:

- Agent → model it calls → key it uses.
- Agent → tools and MCP servers it can reach → what each tool can write.
- Agent → identity it runs as → human it acts on behalf of.
- Agent → version history → who changed it last.

Without these links you can count assets but cannot answer "if this key leaks, which agents break?" or "which agents can send email?"

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud. Connectors reach third-party agent platforms (Salesforce, SAP, Databricks, HubSpot) and Claude Enterprise and OpenAI Enterprise workspaces, and it discovers and manages skills in Anthropic Enterprise. Everything lands in a governed registry of agents, versions, tools and MCP servers, and Posture flags risky permissions and configurations. Opencontroller Edge, installed by IT via MDM, extends this to developer devices, coding agents, skills and local MCP servers. See the [capability reference](/capability-reference) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Should API keys be in the AI inventory or the secrets inventory?**
Both. The secrets manager holds the key; the AI inventory records which agents and people use it, what it costs and when it was last used.

**Do personal skills and custom GPTs count as AI assets?**
Yes, once they are shared or connected to company data. A shared assistant behaves like a small agent and needs an owner.

**How often should the inventory refresh?**
Cloud and SaaS sources can refresh daily or faster. See [how often AI agent discovery should run](/blog/how-often-should-ai-agent-discovery-run) for a cadence by source.

## Key facts

A complete AI inventory covers ten asset types: agents, models and endpoints, API keys, tools and MCP servers, skills, workspace assistants such as custom GPTs, coding agents, prompts and system instructions, vector stores, and agent identities. NIST AI RMF GOVERN 1.6 calls for mechanisms to inventory AI systems. Useful entries link each agent to its model, key, tools, identity and owner. Opencontroller by Lyzr auto-discovers agents, models and API keys across AWS, Azure and Google Cloud, connects to Salesforce, SAP, Databricks, HubSpot, Claude Enterprise and OpenAI Enterprise, discovers skills in Anthropic Enterprise, and with Opencontroller Edge covers coding agents and local MCP servers on developer devices. Pricing starts from $250K per deployment per year.

## Related

- [How to find every AI agent in your enterprise (including shadow agents)](https://opencontroller.lyzr.com/blog/how-to-find-shadow-ai-agents.md)
- [What permissions does AI agent discovery need in AWS, Azure and Google Cloud?](https://opencontroller.lyzr.com/blog/ai-agent-discovery-permissions-aws-azure-google-cloud.md)
- [What is agent skills governance?](https://opencontroller.lyzr.com/blog/what-is-agent-skills-governance.md)
- [More on Discovery and shadow agents](https://opencontroller.lyzr.com/blog/cluster/discovery-and-shadow-agents)
