Connect an external agent

You already run an agent on your own stack and you do not want to rebuild it here. This is how you register it in your Workspace: The platform keeps the management, the deployment channels and the reporting, while the thinking stays on your side.

Open it from the Console left menu, under BUILD, item External Agents.

When this is the right choice

The agent already works on your own stack, for example n8n, LangGraph, an internal service or a partner product.
You want the Workspace, the deployment channels and the reports without rewriting the agent logic.
You want to keep control of your data: The knowledge and the processing stay inside your own infrastructure.

To build an agent on the platform itself, use My agents instead.

What you need before you start

What	Why
Base URL	The root address of your agent, for example `https://agent.acme.com/ai`. The platform calls `/health` and `/runs` under it
The `fpt-v1` contract	Your agent has to speak the platform’s agent protocol. That part belongs to the team that built it
Credentials	The header and value the platform should send, if your system checks callers

The Integration guide button at the top right of the list screen is written for the engineers on the partner side. To fill in this form you only need the Base URL and the credentials they hand you.

Register an agent

Open External Agents

Console left menu, group BUILD.

Press Connect External Agent

The blue button at the top right. On an empty workspace the same button sits in the middle of the screen.

Fill in the connection

Agent name, Description, Base URL and how the platform should authenticate. Every field is described below.

Press Validate Connection

The platform calls your system and scores five checks. Save stays locked until all five pass.

Press Save

The agent is created as a Draft and Console opens the detail page. Nothing is written before you press Save.

The fields

Field	Required	What it is	Example
Agent name	Yes	The name people see in the list and on the channels. Up to 60 characters	Warehouse assistant
Description	No	What the agent does and when to reach for it. Up to 200 characters	Looks up stock levels and order status
Base URL	Yes	The root address of the agent. Both https and http are accepted	`https://agent.acme.com/ai`
Authentication	Yes	Either Authenticated or No authentication	Authenticated
Auth headers	Yes, when Authenticated	Header name and value. Add more with Add header	`Authorization` / `Bearer abc123`

Base URL accepts http, but then the credential travels in clear text. Keep http for an internal partner and use https everywhere else.

The two authentication modes

Authenticated: The platform sends the headers exactly as you typed them and adds nothing of its own. For a bearer token, type the word Bearer in front of the value yourself. The value is stored and never shown again.
No authentication: The platform sends nothing. Pick this only when the partner already restricts access another way, for example by IP range.

The five checks

Check	What it proves
Endpoint reachable	Your agent answered a health check
Partner accepted our credentials	Your system accepted what the platform sent
Partner supports `fpt-v1`	Your agent speaks the platform’s protocol
Credential stored	The credential is saved. Publishing waits on this
The agent answers a real turn	The platform sent one turn to `/runs` and your agent replied in the contract. A health check alone cannot tell you this

Under the list the partner reports its own name and version, together with the round-trip time. A check that fails explains why, so you can correct the form and press Validate Connection again.

The list of connected agents

Each agent is a card with its name, status, description and when it changed last. The filters above count the agents in each state: All, Draft, Published, Paused. The Search agents box on the right filters by name.

Inside one agent

The detail page has four tabs: Build, Test, Channels and Insights. The Build tab is where the connection lives:

Row	What it tells you
Base URL	The address in use
Authentication	Authenticated or not
Credential	Stored, or not needed
Health	The result of the last check, with a Check again button

The Endpoints table lists the addresses the platform calls on your agent and the state of each one. /health and /runs are required; the rest are optional and off in this phase.

After the connection is live

From here an external agent behaves like any agent built on the platform:

The Test tab sends a real turn so you can read the answer before anyone else does.
The Channels tab picks the Workspace audience and the outside channels: Web widget, API, Zalo.
Publish at the top right puts a version into service.

These three work exactly as they do for a normal agent, so see Publishing and Managing deployment channels.

Edit, recheck and delete

Edit the connection

On the detail page, press the pencil next to Publish, or the Edit connection button. The dialog is the same one you used to create the agent and it runs the same checks before it saves.

Run the checks again

Check again on the Connection card, or Recheck now on the Endpoints table. Run check now in the … menu does the same thing, both on the detail page and on the card in the list.

Delete the agent

The … menu, then Delete agent. Console asks you to confirm.

A new Base URL or credential only reaches real users after you publish again. Deleting the agent pulls it off every live channel, and customers in the middle of a conversation stop getting answers.

Next steps

Managing deployment channels

Pick the channels and the audience for an agent.

Live Chat channel

Put the agent on your website with the ready-made chat window.

API integration

For when you already have your own chat interface.

Introduction

FPT AI Console

Connect an external agent

When this is the right choice

What you need before you start