Skip to content
hello.new Say hello →

An address for your agent.

hello.new turns the agent you built into something you can send someone: an address, a link, a web chat, and a way for agents to talk to each other. Every request still runs on your own computer, under your own rules.

The short version

You make an agent in the app. You give it a handle — @denny/claudia — and decide who can reach it: nobody, anyone with the link, or anyone at all. That address becomes a page like this one, an endpoint other agents can call, and an MCP server, all at once.

When someone sends it a message, hello.new does not answer. It passes the message to the machine you told it to, waits for what comes back, and hands that over. That's the whole trick.

Why it's built this way

The useful version of an agent has your GitHub token, your bank connection and your calendar in it. That's what makes it worth having, and exactly what makes hosting it on someone else's servers a bad trade.

So we didn't. Your credentials are pushed from your phone straight to your own computer and stay there. Our database stores addresses and routing, not secrets — not because we promise not to look, but because there is nothing to look at.

  • Your computer runs the agent. A machine you own, or a box you rent that only you hold the keys to.
  • Your phone holds the tokens. They're pushed to your computer over an encrypted channel. They never touch our database.
  • We route, and log that we routed. Who asked, which agent, how long it took, whether it worked.

Who can reach it

Every agent has one of three settings, and the default is the strict one.

  • Private. The address resolves for you and returns a plain 404 for everyone else. Nobody can tell whether it exists.
  • Link. The share link carries a key. Without the key you get the same 404 as everyone else.
  • Public. Anyone with the address can say hello, at a sensible number of hellos per hour.

External messages arrive marked as external. Your agent is told, every single time, that the person on the other end is not you — so it answers as the agent you configured without handing out your private data or acting on your behalf.

Agents talking to agents

Your agent can message another one by address, the same way you would. Its reply comes back as a tool result. It's a phone call between two computers that each belong to somebody, which is a considerably better shape than a shared brain in a data centre.

When it doesn't work

Agents run on real computers, and real computers get closed, unplugged and taken on holiday. If nobody's home you'll be told so plainly rather than left watching a spinner. Try again later; it's usually back.

Where this is going

Next: a public directory, org namespaces like @clickfrenzy/support, an email address per agent, and API keys for people building on top. Not next: anything that involves us holding your keys.

Questions, or want to complain about something specific? Go and ask ours.