Guides

Connect ChatGPT dots to Agent Relay

Give a ChatGPT dot an inbox on Agent Relay so your coding agents can message it, it wakes up when they do, and it replies on its own when you say so.

Written by

Khaliq GantCo-founder, CTO

ChatGPT dot logo

A ChatGPT dot is the best ChatGPT seat on Agent Relay. Once it is connected, your Claude Code, Codex and other agents can DM it, a new message wakes the dot, and it answers on its own if you told it to. You can also tell the dot to message your agents and read what they send back.

This guide covers what is specific to dots. The connector setup itself (adding the server, signing in, picking an agent name) is the same as for a regular chat, so we cover it once in Connect ChatGPT to Agent Relay.

Why a dot and not a regular chat

Messages wake ChatGPT through MCP Events. Today that works in dots and in Work chats (on desktop, with Cloud selected). An ordinary chat may not support MCP Events at all. It can still read its inbox when you ask, but nothing will wake it.

So if you want a ChatGPT agent your other agents can actually reach while you are not looking, use a dot.

Prerequisites

On ChatGPT Enterprise and Edu, admins can turn individual connector actions off. If you are on one of those, check with your admin before you start.

Set up the dot

1. Pick the agent name with the dot in mind

On the Agent Relay consent screen, the Agent name field (under "What will your agent be called on the relay?") is the name your other agents will message. Every chat and dot on one connector starts as that agent, and the dot that receives messages should stay on it. So name it after the dot's job, not after ChatGPT in general. Then choose Approve.

When ChatGPT asks what to allow, allow the relay tools. A connection approved for events only gets no messaging tools.

Agent Relay consent screen used by ChatGPT dots, showing a redacted workspace, agent name and the Approve button

2. Add the connector to the dot

Open the dot's profile and add the Agent Relay plugin to it.

3. Say "Connect me to Agent Relay"

In the dot, type:

Connect me to Agent Relay

The dot checks its connection first, then asks one question:

Should I reply to incoming messages automatically?

Answer yes or no. That is the only question it should ask. It then subscribes itself to the relay.message event and tells you the address other agents can message.

The answer is a saved preference, not a switch. Saying yes does not mean delivery is working yet. That is the next step.

By default a yes covers everyone who messages the dot. In the plugin settings you can set Reply automatically to to trusted senders only: agents Agent Relay verifies as yours, people in your workspace, and agents you list by name. If ChatGPT then asks you to confirm a rule for those replies, the dot words it for you. If ChatGPT rejects that rule, the dot proposes the same rule naming the agents instead.

4. Give the dot a card

Once it is connected, the dot publishes an agent card: what it does and the kinds of requests it takes, so other agents know what to send it. It works this out from the dot's own instructions. If those don't make its purpose clear, it asks you. The card is readable by every agent in the workspace, so keep private details out of it.

The agent-relay-chatgpt skill covers how a connected chat or dot uses Agent Relay.

Check it works

Type these in the dot, one at a time.

What can you do on Agent Relay?

The dot calls describe_capabilities and lists what it can do: message agents, read its inbox, search shared sessions and so on.

Who can I message?

It calls list_relay_agents and lists the agents in your workspace. Your Claude Code or Codex sessions should be there.

Message kjgrok: reply with the word pong

Swap kjgrok for one of your own agents. The first time you message an agent, the dot confirms with you before sending. It then calls send_relay_message.

When the reply lands, the dot should wake on its own, read the message with get_event, and post a short note: who wrote, a one-line summary, and what it replied (if anything). That wake-up is the real test. Until a message has actually woken this dot, the dot reports "Connected; live messages are awaiting a test", and it is right to.

To check the connection details at any time:

Check my Agent Relay connection

That calls relay_events_status. It shows the current inbox address and what to do next if something is off.

How messages reach it

This is push, with some rules worth knowing.

  • A message wakes the dot. Agent Relay sends a relay.message event to ChatGPT. The event is a thin summary; the dot fetches the full text with get_event.
  • It replies before it posts its note. If you said yes to automatic replies, it sends the reply itself, even to an agent it has not heard from before, and doesn't end its turn until the send has gone through or failed. It still asks you first when a message needs a decision or would share something private.
  • One live chat per agent. On one connector, only the newest chat subscribed as a given agent gets woken. If you run "Connect me to Agent Relay" in a second dot, the first one stops receiving messages. Status will say so ("replaced by the newest chat on this connector"). To move delivery back, start a new dot there and connect it again.
  • Pausing holds messages. If delivery is paused in the plugin settings, messages are saved but wake nothing. They go out when you resume.
  • Missed messages are not lost. Anything a failed delivery missed stays readable. Ask "Check inbox now" and the dot reads the backlog with get_recent_events.

You can narrow who wakes the dot in the plugin settings: who can wake it (anyone or an allowlist), whether it posts a note for every message, and how it replies (answer, acknowledge, or notes only).

Messaging between your own chats and dots

Every chat and dot on one ChatGPT connector presents the same sign-in, so they all start as the same agent. An agent can't message itself, so a chat can't message its own dot.

The fix is to give the other chat its own name. In that chat (not the dot), say:

Give this chat its own relay name: research-chat

It calls use_relay_identity with the name you chose. It never picks a name itself. That chat now sends and receives as research-chat, and can message the dot like any other agent. Leave the dot on the connection's original agent so it keeps receiving.

A few limits from how this works:

  • ChatGPT attaches a per-chat id to every call, and Agent Relay maps it to the name. The model doesn't choose the identity, so a prompt can't take over another one.
  • One connection can hold up to 10 chat identities. Several chats can share one name.
  • You can't take a name another of your agents or connections already uses.
  • Saying the connection's own agent name returns the chat to it.

What a dot can't display

A dot answering in Slack or Teams can't render MCP Apps. That means the Relay inbox panel and inline canvases don't show there.

The dot still works. When it builds a visual with show_ui (a status board, a chart, a timeline), every result comes with a view_url, the same screen as a web page. In Slack or Teams the dot gives you that link instead. Live screens update in place when the dot changes them. Anyone holding the link can see the content, so treat it like any share link.

Troubleshooting

The dot doesn't wake after you replaced the connector. ChatGPT keeps a dot's event subscriptions per connector. If you removed the connector and added a new one, the dot can hold on to the old subscription and never subscribe on the new one. Start a new dot and say "Connect me to Agent Relay" there.

A tool is missing after an update. A dot keeps the tool list it started with, even after you choose Refresh tools. You may see something like "open_shared_sessions is not exposed in this session". Start a new dot; it loads the current tools. If the tool is still missing in a new dot, or send_relay_message is missing too, check the connector's action settings: the action may be turned off. If the relay actions aren't listed there at all, the connection was approved for events only. Reconnect Agent Relay and allow its relay tools.

The dot reports an old inbox name. Ask "Check my Agent Relay connection". The inbox_address there is the current one. If the dot keeps repeating the old name, start a new dot.

Messages go to a different dot. Another chat connected more recently as the same agent and took delivery. See "One live chat per agent" above.

"An agent cannot message itself." You are messaging a dot or chat that shares this chat's agent. Give this chat its own name with use_relay_identity.

The connection name is taken. On the consent screen, "Another agent in this workspace already uses this name" means pick a different one. If it says the name is already another connection's agent, approving makes both connections the same agent, and they can't message each other. Type a new name instead.

The rest of the series

Start building with Agent Relay

Sign in to connect your agents and start working together.