---
title: What is an AI agent registry?
url: https://opencontroller.lyzr.com/blog/what-is-an-ai-agent-registry
published: 2026-11-13
updated: 2026-11-13
author: Lyzr
---

# What is an AI agent registry?

**An AI agent registry is the system of record for every AI agent in an organisation: what it is, who owns it, which version is running, which models, tools and MCP servers it uses, what it is allowed to do and what risk tier it sits in. It is where an agent becomes governed, as opposed to merely discovered.**

Buyers often ask whether their CMDB or their model registry can do this job. Both hold part of the picture. Neither was designed for things that act on your behalf.

## How is an agent registry different from a CMDB or a model registry?

| Question | CMDB | Model registry | Agent registry |
|---|---|---|---|
| Main object | Servers, apps, services | Trained model artefacts | Agents and their versions |
| Typical owner | IT operations | Data science / ML platform | Platform, security and risk jointly |
| Tracks tools and MCP servers an agent calls | No | No | Yes |
| Tracks the identity the agent acts as | Sometimes, as a service account | No | Yes, including on-behalf-of delegation |
| Tracks which models an agent calls at runtime | No | Lists models, not their callers | Yes |
| Holds evaluation results per version | No | Sometimes, for models | Yes, for the agent as a whole |
| Holds risk tier and approval status | Rarely | Sometimes | Yes |
| Updated by | Discovery tools and change tickets | Training pipelines | Discovery, CI/CD and connectors to agent platforms |

A CMDB tells you a container runs on a cluster. A model registry tells you which model version was trained and approved. Neither tells you that a procurement agent, version 3, calls two external models and a supplier-portal MCP server while acting on behalf of a named buyer. That is the agent registry's job.

In practice, the agent registry links to both: the CMDB entry for where it runs, and the model registry entry for any in-house model it calls.

## What records should an agent registry hold?

At minimum, each agent entry needs:

1. **Identity:** a stable agent ID, name, description and the platform it runs on.
2. **Ownership:** a named business owner and technical owner, resolved against your identity provider.
3. **Versions:** each deployed version, with its prompt or configuration hash and release date.
4. **Dependencies:** models, tools, MCP servers, skills, knowledge sources and other agents it calls.
5. **Access:** the credentials or workload identity it uses, and whether it acts as itself or on behalf of a user.
6. **Risk:** risk tier, data classes it touches, and the approvals it received.
7. **Evidence:** links to evaluation runs, red-team results and policy decisions per version.
8. **Status:** proposed, approved, production, quarantined or retired.

Our [AI agent inventory fields guide](/blog/ai-agent-inventory-fields) goes field by field, and the [MCP server inventory fields guide](/blog/mcp-server-inventory-fields) covers tool records.

## How do agents get into the registry?

Through three routes, and you need all three:

- **Discovery.** Scanning cloud accounts and agent platforms finds agents nobody registered. See [how to find shadow AI agents](/blog/how-to-find-shadow-ai-agents).
- **Self-registration.** Teams register agents through a form or a CI/CD step before release.
- **Connectors.** Integrations with platforms that host agents keep entries in sync as agents change.

A registry fed only by self-registration goes stale. A registry fed only by discovery lacks owners and intent. Combine them, and reconcile differences weekly.

## Does registering an agent mean controlling it?

No, and this is a common source of confusion. Registration gives you a record. Control depends on where you can act:

1. **Registry:** you know the agent exists and who owns it.
2. **Observe:** you see its model calls, tool calls or platform events.
3. **Enforce:** a gateway or platform control can allow, route, redact or block its actions.

Many agents start at level 1 and move up as traffic is routed through gateways or platform connectors expose controls. Track the level per agent so nobody assumes coverage you do not have.

## How Opencontroller does this

Opencontroller provides a governed registry of agents, versions, tools and MCP servers. It is fed by auto-discovery of agents, models and API keys across AWS, Azure and Google Cloud, and by connectors to Salesforce, SAP, Databricks, HubSpot and Claude Enterprise and OpenAI Enterprise workspaces. Owners resolve through OIDC SSO, and agents can carry delegated on-behalf-of identity or federated workload identity. Registration does not mean runtime takeover: depth of observation and enforcement depends on the connector and traffic path. Read [what an AI agent control plane is](/blog/what-is-an-ai-agent-control-plane) for the full picture, or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Can we use our CMDB as an agent registry?**
You can store some agent records there, but most CMDBs lack fields for tools, MCP servers, delegated identity, evaluation evidence and runtime model calls. Most teams keep a dedicated agent registry and link to the CMDB.

**Who should own the agent registry?**
Usually the platform or AI enablement team runs it, security and risk set the required fields and approvals, and each agent's owner keeps their entry accurate.

**How often should the registry be reconciled with discovery?**
At least weekly, and on every production release. Any agent found by discovery but missing from the registry should be treated as a shadow agent until an owner claims it.

## Key facts

An AI agent registry is the system of record for AI agents: identity, owners, versions, models, tools and MCP servers used, identity and delegation, risk tier, evaluation evidence and status. It differs from a CMDB (infrastructure) and a model registry (trained model artefacts). Registration is not runtime control; control depth goes registry, observe, enforce. Opencontroller by Lyzr provides a governed registry of agents, versions, tools and MCP servers fed by auto-discovery across AWS, Azure and Google Cloud and platform connectors. Pricing starts from $250K per deployment per year.

## Related

- [What fields belong in an AI agent inventory?](https://opencontroller.lyzr.com/blog/ai-agent-inventory-fields.md)
- [What should you record for every MCP server in your inventory?](https://opencontroller.lyzr.com/blog/mcp-server-inventory-fields.md)
- [How to find every AI agent in your enterprise (including shadow agents)](https://opencontroller.lyzr.com/blog/how-to-find-shadow-ai-agents.md)
- [More on Agent governance basics](https://opencontroller.lyzr.com/blog/cluster/agent-governance-basics)
