Hermes and OpenClaw bots on OpenClaw Launch can now be added to ChatGPT or Claude as a connector, so the assistant you already type into can ask the one that actually knows your context.
You know the shape of this already. You're in ChatGPT, halfway through something, and you need a fact that isn't in the chat. What's on your calendar tomorrow. What you decided in the notes you've been keeping. Whether that thing you asked for on Monday ever got done.
You have a bot that knows. It's been running for weeks. If you connected your calendar to it, it can read your calendar. It wrote the notes. It has the skill you set up for exactly this kind of question. But it lives in another tab, and the only bridge between the two is you, copying a question across and pasting an answer back.
That bridge is now a setting. Copy one address, approve it once, and you can ask ChatGPT to ask your bot. The answer comes back in the same chat.
What a connector actually is
The feature is called an MCP connector. MCP is a standard that lets an AI app call something outside itself, and you don't need the acronym to use it. Think of a plug socket. ChatGPT gets one socket that leads to your bot.
The part worth understanding is what doesn't happen. Your bot isn't copied into ChatGPT. Its memory, its skills and the apps you connected to it stay where they are, running on your bot. ChatGPT gets a way to send it a message and read what comes back. That's the whole mechanism.
This matters because the alternative is worse than it sounds. Pasting an answer into ChatGPT gives it a paragraph. It doesn't give it the twenty previous conversations that made the answer make sense, or the notes file the answer came from, or the ability to go and check something else tomorrow. A connector keeps all of that on the bot's side, where it already lives.
What it feels like
You type the way you'd text a person who happens to know your stuff.
Ask my bot what's on my calendar tomorrow.
ChatGPT hands the message to your bot. Your bot works the way it does in Telegram or WhatsApp, with its own model, its own memory and its own connected apps, and the reply appears in the ChatGPT conversation a moment later. You stay where you were.
Some other things this is good for:
- "Have my bot add this to the project notes it keeps."
- "Ask my bot to run its weekly report skill now."
- "Ask my bot whether the thing I mentioned this morning is finished."
Follow-up questions continue the same conversation on the bot's side, so it still has the earlier context when you ask the second question. If you'd rather start clean, tell ChatGPT to start a new conversation with your bot first.
Why there's only one tool
The connector gives ChatGPT exactly one ability, and it's called ask_bot. Send a message, get the reply. There is no file tool, no terminal, no way to change your bot's configuration or read its credentials.
That limit is deliberate. A connected app should never be able to do more than a person typing to the bot could do, and that's the line this sits on. If you wouldn't hand a stranger your Telegram bot, don't hand them the connector address either.
Access is the part I'd want you to read properly, because this does give an outside app a route to an agent with real permissions. When you add the connector, ChatGPT sends you to an OpenClaw Launch page that names the exact bot being connected. You sign in with the account that owns it and press Allow. What the app receives is a token that works for that one bot. Your bot's API keys and provider credentials never leave the platform, and the connector address on its own does nothing until someone signs in and approves it.
If you change your mind, open the bot's MCP window and press Disconnect. Access ends there and then, not at the next restart or the next billing date. You can also remove the plugin on the ChatGPT side, though disconnecting on the dashboard is what actually revokes it.
One warning that applies to any connector, not just this one: only approve a connection you started yourself. If someone sends you a link and asks you to approve it, close the tab. Whoever completes that screen can use the bot, and everything connected to that bot.
Setting it up
Two things before you start. You need a running bot on OpenClaw Launch with an active plan or free trial, and you need a ChatGPT plan that allows custom connectors. ChatGPT Free doesn't offer them, and on a work workspace an admin may have switched them off. Claude accepts custom connectors on any plan, with one connector on Free.
For ChatGPT:
- On your dashboard, press MCP on the bot's card and copy the Connector URL from the ChatGPT / Claude tab.
- Open ChatGPT's plugins page, press Add and choose Add custom MCP server.
- Give it a name you'll recognise, paste the URL under Connection, and leave Authentication on OAuth.
- Tick the confirmation, press Create as a plugin, then Continue.
- Sign in to OpenClaw Launch, check that the bot named on the page is the one you meant, and press Allow.
- Start a new chat and try: "Use my bot to tell me what you can do."
For Claude:
- Copy the same Connector URL from the bot's MCP window.
- Go to Customize, then Connectors, press + Add and choose Add custom connector.
- Paste the URL, add it, and press Connect.
- Sign in, check the bot name, and press Allow.
On Claude Team and Enterprise, an owner adds the connector first under Organization settings, then Connectors, and each member connects it for themselves.
Those menu names were correct when the flow was checked in October 2026 on a Pro plan. Both companies move their menus around, so if a button has a different name, look for the option to add a custom connector rather than for the exact wording above.
If ChatGPT answers the question itself instead of passing it to your bot, name the plugin in your message. "Use the your bot's name plugin to ask my bot what's on my calendar tomorrow" is usually enough, and after that it tends to get the idea.
What it doesn't do
It doesn't make ChatGPT smarter, and it doesn't give ChatGPT your bot's model. ChatGPT relays the message and reads the reply. The work happens on the bot, which spends the same AI credits as a message from any other chat app.
A single request can run for just under five minutes. For anything longer than that, ask the bot to start the job and message you when it's finished, rather than leaving the ChatGPT request hanging in the meantime.
It's also not a way to reach a bot that isn't yours. Bots created through the developer API, and bots you resell to your own customers, can't be connected this way.
If you don't have a bot yet
None of this is much use without one. OpenClaw Launch runs Hermes or OpenClaw for you, with restarts, updates and chat apps handled from a dashboard, and the connector is included on every plan and during the free trial. A trial takes a minute to start and doesn't ask for a card.
Top comments (0)