---
title: How do you register an AI agent your team doesn't host?
url: https://opencontroller.lyzr.com/blog/register-ai-agents-you-dont-host
published: 2026-11-12
updated: 2026-11-12
author: Lyzr
---

# How do you register an AI agent your team doesn't host?

**Register it by what you can reach: its platform's admin API, its endpoint, or the traffic it sends through your gateway. Record a minimum entry (owner, purpose, data, tools, identity, endpoint) on day one, then add observation and enforcement as far as the traffic path allows. Registration does not require you to host, rebuild or migrate the agent.**

Many agents in an enterprise are not hosted by the team that answers for them. A vendor runs one in its SaaS product. A business unit built one on a cloud platform. A partner exposes one behind an API. You still need it in the inventory.

## Why does an agent you don't host still need registering?

Because you are still accountable for what it does with your data and your users. Regulators and auditors ask the deployer, not only the vendor. An agent outside the inventory is a [shadow agent](/blog/how-to-find-shadow-ai-agents), even if procurement approved it.

Registration answers four questions: what it is, who owns it, what it can touch, and how you would stop it.

## What are the ways to register an external agent?

There are four routes. Pick the deepest one you can reach.

| Route | How it works | What you can observe | What you can enforce |
|---|---|---|---|
| Manual record | Owner fills in a registration form | Nothing live; only what is declared | Nothing at runtime; review and approval only |
| Platform connector | Read the agent's config and activity through the platform's admin API | Configuration, versions, permissions, often usage | What the platform's API allows, such as disable or restrict |
| Endpoint registration | Record the agent's API endpoint and call it through a governed client | Your calls to it, inputs and outputs | Who may call it, and policies on what you send and receive |
| Traffic path (proxy or gateway) | Route the agent's model or tool calls through your gateway | Every model and tool call that passes through | Allow, route, redact or block in the path |

The deeper the route, the more you see. A manual record is better than nothing, but it is a promise, not a control. Read more on the trade-offs in [read-only vs inline governance](/blog/read-only-vs-inline-ai-agent-governance).

## What is the minimum record for an agent you don't host?

Capture these fields at registration. Leave nothing blank; write "unknown" and set a date to find out.

1. **Owner** — a named person in your organisation, and a backup.
2. **Operator** — who hosts and runs it (vendor, business unit, partner).
3. **Purpose and scope** — what it is for and what it must not do.
4. **Data classes** — what it reads and writes, and where data is processed.
5. **Tools and actions** — systems it can call, including MCP servers.
6. **Identity** — whose credentials it uses: a service account, delegated user access or the vendor's.
7. **Endpoint or platform** — where it lives and how to reach it.
8. **Kill path** — how to disable it and who can do it.
9. **Route** — which of the four routes above you used.
10. **Review date** — when to re-check the record.

Use the [registration questionnaire](/blog/ai-agent-registration-questionnaire) for the full form.

## How do you move from a record to real control?

Step through it over time:

1. **Register** with the minimum record.
2. **Connect** to the platform's admin API if one exists, so config and versions update themselves.
3. **Observe** by sending traffic through your gateway where the agent allows a configurable model or tool endpoint. Many frameworks accept an OpenAI-compatible base URL.
4. **Evaluate** a sample of real interactions for groundedness and task adherence.
5. **Enforce** policies in the path, and agree a kill path with the operator for when the path is not yours.

For vendor-run agents with no API and no configurable endpoint, contract terms carry the weight: logging access, notice of model changes, and a right to suspend. See [governing agents built by vendors](/blog/govern-ai-agents-built-by-vendors).

## How Opencontroller does this

Opencontroller, Lyzr's open control plane for enterprise AI agents, keeps a governed registry of agents, versions, tools and MCP servers, whether discovered automatically across AWS, Azure and Google Cloud, pulled through connectors (Salesforce, SAP, Databricks, HubSpot, Claude Enterprise and OpenAI Enterprise workspaces) or registered directly. Registration is not runtime takeover: the depth moves from registry to observe to enforce depending on the connector and traffic path. Agents that route through the LLM Gateway or Agent Gateway get policies in the path, with control and no migration. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Can I govern a vendor's agent without access to its code?**
Partly. You can register it, govern your calls to it and its data access, and use contract terms. Runtime enforcement needs a connector or a traffic path.

**Who should own an agent a vendor runs for us?**
A named person in your organisation who answers for its use, usually the business owner of the process it serves, with a technical backup.

**Is a manual registration enough for an audit?**
It shows the agent is known and approved, but not how it behaves. Auditors increasingly expect live evidence, so add a connector or traffic path where you can.

## Key facts

An AI agent your team does not host can be registered four ways, in increasing depth: a manual record, a platform connector (admin API), endpoint registration, or routing its model and tool traffic through a gateway. The minimum record has ten fields: owner, operator, purpose and scope, data classes, tools and actions, identity, endpoint or platform, kill path, route and review date. Only connectors and traffic paths give live observation; only the traffic path allows allow, route, redact or block in the path. Opencontroller by Lyzr registers agents discovered across AWS, Azure and Google Cloud, through connectors, or directly; registration is not runtime takeover. 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 should an AI agent registration questionnaire ask?](https://opencontroller.lyzr.com/blog/ai-agent-registration-questionnaire.md)
- [How do you govern AI agents built by vendors and system integrators?](https://opencontroller.lyzr.com/blog/govern-ai-agents-built-by-vendors.md)
- [More on Discovery and shadow agents](https://opencontroller.lyzr.com/blog/cluster/discovery-and-shadow-agents)
