Huddles, with your agent in the room

Here is a thing that happened last week. I was deep in a terminal, three files open, and a thought arrived that had nothing to do with any of them: I should talk to Chad about the draft. The old move is to stop, find the tab, find Chad, press the thing. The new move is to not stop. I typed start a huddle with Chad into the agent I already had open, kept reading my diff, and twenty seconds later Chad was in my ear saying "you rang?"

That is the whole feature, honestly. Huddles now speaks the Model Context Protocol — MCP, the thing Claude Code, Cursor, Windsurf and most of the agents you might keep open all day already understand. You connect it once, with a token you make on one page, and from then on your agent can do in Huddles what you would do with a mouse. Not more. Not for anyone else. Just you, in your spaces.

What it is like

It is less dramatic than it sounds, which is the point. Nothing new to learn; you just say things.

"Start a huddle with Chad." The agent looks at your friends, finds Chad, and asks him in. The huddle opens where you are standing in the space, so you do need the space open — if you do not, the agent hands you the link and tries again once you are there. Chad gets a notice if he is around and an email either way.

"What is on tonight?" The space's calendar, read back to you: what is live right now, what is coming, whether you already said yes.

"Create an event for tonight at 8 on the main space, call it Draft review." The agent works out what "tonight at 8" means in your timezone, puts it on the calendar with you as host, and everyone in the space sees it. People who say yes get the reminder.

"How much have I huddled this week?" Your own numbers — huddles, minutes, this week against last, events you hosted, who you end up talking to most. Yours and nobody else's, which is a rule we will come back to.

"Invite ana@example.com into my space." For a space you own: a link that admits her and nobody else.

None of this needed a new space, a new account, or a new anything. You already have a space. The agent acts in it.

The one-time part

The token lives where every Huddles credential lives: the developer hub. Open it, go to Agents, press Create a token, give it a name like "Claude Code on my laptop". The hub shows the token once and fills in the config for you — a one-line command for Claude Code, a block of JSON for everything else. Paste, done. If a laptop grows legs, revoke the token on the same screen and the agent is deaf to Huddles from that second.

The hub opens by request — a person reads a line about what you are up to and says yes, usually within a day. That is the one gate on all of this, and for now it is deliberate: a token acts as you, and we would rather know who is handing one to a machine. Ask once at huddles.space/developers/access; everything after is a minute's work.

A token acts as you, so it needs an account behind it, and it only reaches the handful of tools on that page. There is no tool for other people's stats, no roster, no moderation, nothing an operator would use. Every tool runs the same route the app itself runs when you click, under the same gates, so there is no second permission system quietly growing in a corner.

The other thing it can do

Some of you are not just using Huddles; you are building something and want Huddles inside it. A course page where classmates can see each other. A messenger that could use a real call button. That is a different ask, and we kept it apart on purpose, because the first thing it needs is a space of its own for the app to open into — and nobody using Huddles day to day should have an agent making spaces for them.

For the builders, the agent can do the setup end to end. Say "add Huddles to my app" (the agents page has a fuller prompt you can paste, with the blanks to fill). It finds out who you are, makes a space for the app if you do not already have one, asks for developer access with a line about what you are building, and then tries to create the app.

And here it stops, and this is the part we are most pleased with. Access to app ids and secrets is given by a person, by email, and no agent can hurry that. So instead of a 403 your agent has to guess about, it gets a structured answer that says: pending, here is what to tell your user. Your agent will say something like "I have set up the space and asked for access — let me know when you are approved and I will create the app and wire the install." Which is exactly the sentence you would want from a good colleague. When the approval lands, you say "go", and it does: the app id, the secret (once, into your server's environment, never into page code), the install for a page or for a huddle button, and the documentation read as text so it can wire your actual code.

Why MCP and not a chat bot

We could have put a chat box in the app that understood "start a huddle with Chad". We did not, for a reason that matters more as agents get good: the agent you already have knows things about you that ours never will. It knows which project you are on, who Chad is in your codebase, that "tonight" means after the kids are down. A protocol lets it bring all of that; a bot in our corner would have to be told from scratch every time.

It also keeps us honest. MCP tools are small, named, and listed. You can read the whole set on the agents page in a minute. There is no "the AI can do stuff" fog — there are fifteen verbs, and each one is a thing you could already do.

Where it goes next

A couple of things are obvious from here and not built yet. Agents should be able to hear Huddles as well as speak to it — a huddle starting, a friend arriving — so your agent can tell you instead of you asking. And the pasted token should become a proper sign-in, so connecting an agent is a button rather than a copy and paste. Both are on the list. For now: one page, one token, and your agent in the room.

If you try it, tell us what you said to it. The sentences people actually use are the best spec we have.