ChatGPT dots

A dot is ChatGPT’s always-on agent. It runs in the cloud and can use the plugins installed and enabled for your account. HALO becomes such a plugin when you add it as a custom MCP server.

1. Give the server a public address

A dot runs in OpenAI’s cloud, so it cannot reach localhost. Start the plugin behind a tunnel. With a Cloudflare quick tunnel, which needs no account:

cloudflared tunnel --url http://127.0.0.1:8765

Note the https://<name>.trycloudflare.com address it prints, then start HALO with that host name:

halo serve --http --port 8765 --public-host <name>.trycloudflare.com

HALO plugin URL: https://<name>.trycloudflare.com/<random>/mcp

The random path segment keeps the server private to whoever has the full address. Every other path returns 404. A quick tunnel gets a new name each time it starts, so the plugin’s address changes with it. A named tunnel or your own HTTPS host keeps it fixed.

2. Add HALO as a custom MCP server

1. Open chatgpt.com/plugins, choose Add, then Add custom MCP server.
2. Name it HALO, paste the plugin URL, and set authentication to No authentication.
3. Read the risk notice, confirm it, and choose Create as plugin. ChatGPT connects to the server and lists its tools.
4. In the Connect HALO dialog, choose Connect. The dialog says that ChatGPT may share relevant memories and chat content with the plugin.

HALO now appears under Personal on the plugins page.

3. Ask the dot to use it

Open the dot’s conversation, type @ and pick HALO, then give it the agent instructions. A dot that has the instructions in its memory keeps logging in later conversations and in the actions it starts by itself.

A first test

On 31 October 2026 we connected HALO to a dot that sends a daily countdown to a conference deadline, a reminder the user had set up. We asked the dot to log the most recent run of that reminder and to call halo_report. It logged seven events and marked each as logged afterwards, with the actual time:

Id	Kind	Step	Uses
e1	observe	the scheduled countdown run starts
e2	observe	reads the official dates	e1
e3	recall	looks back for new feedback from the user, finds none	e1
e4	act	computes the time left and prepares suggestions	e1, e2, e3
e5	act	sends the countdown	e1 to e4
e6	outcome	the message tool accepts the send, about e5	e5
e7	remember	keeps the dates, the time of the reminder, and the send record	e2, e4, e5, e6

The report placed the trace at L0. It observed S→A, S→M, M→A, A→S (e6 checks e5), and A→M (e7 keeps e5), and the record and outcome steps. It did not observe M→S, attribution, or a revision. The dot explained why the trace stays low. Everything happened in one session, so nothing shows memory used in a later session, which L1 needs. The reminder was set up by the user, so both actions are logged with trigger request, which rules out L3 as well. That is the reading the survey gives a reminder the user scheduled.

A trace from a few days of ordinary use, with the instructions in the dot’s memory, is the next step.