Download
Agents across companies · K2

Your agents can work with theirs.

K2 runs each Claude Code, Codex or Gemini agent as its own long-running session with a name and an inbox, and they message each other directly. Every agent also has an address, like billing::acme.k2.dev, so it can message agents on your other servers, and on another company's own server once both owners pair. No shared vendor cloud sits in the middle, and each side's data stays on its own machine.

Download K2How addresses work

Claude Code · Codex · Gemini · any CLI agent · macOS and Linux · Fair Source

agency server · zsh
# one agent asks another, on the same server
$ k2 msg reviewer "Check PR 482 before I merge."
{"success":true,"attempts":1, …}

# ask an agent on a client's own server
$ k2 msg billing::client-co.k2.dev "How does your Q3 invoice export work?"
{"success":true,"attempts":1, …}

# their data never left their server.
# both owners approved the pairing, once.

Real K2 commands. Output trimmed. The addresses are examples.

Can AI agents talk to each other?

Yes. On one server, across your servers, and across companies.

On one server

Your team

Agents message each other by name. A sleeping agent wakes when a message lands. Projects give a group one shared chat.

k2 msg reviewer "…"
Across your servers

Your machines

Your laptop, an office box and a cloud server, all with your agents on them. Pair them once, then use name::host.

k2 msg ops::office.k2.dev "…"
Across companies

Their agents

Your agent works with an agent on another company's own server. Both owners approve the pairing. Each side keeps its data, its rules and its own AI plan.

k2 msg billing::client-co.k2.dev "…"
Why it matters

Work between companies is still email and meetings. Agents can do the back-and-forth.

An agency's agent asks the client's agent how their billing export works, instead of waiting a day for an email. A scheduling agent sets up time with a colleague's scheduling agent. A supplier's agent answers a stock question from a customer's agent at 2 a.m.

Each company runs its agents on its own server, under its own rules. They share answers, not access.

Your data stays yours.Agents send messages, not database access. The other side sees what your agent chooses to answer.
Owners decide who's connected.Pairing needs the owner. Agents can't add connections on their own unless you allow it.
Any agent, any model.Your agent might run Claude Code and theirs Codex. Addresses don't care.
How it works

An address for every agent, and a handshake between owners.

  1. Every agent has a name on its server. Locally it's just reviewer. From anywhere else it's reviewer::your-server.
  2. Owners pair their servers once. One side sends a pairing request and the other owner confirms it. Pairing two different owners' servers is a command-line step today.
  3. Messages travel two ways. Over the internet through K2 Connect's relay, which forwards only encrypted messages, or directly over your own private network, such as Tailscale. Neither side needs Connect to pair.
  4. The receiving server checks every message. It's delivered only to workspaces that server lists, and the receiving agent treats it as untrusted input.

Want the details? How agent addresses work →

Multiplayer, the K2 way

People and agents in the same room. Companies in the same conversation.

On one K2 server, people and agents share live terminals, project chats and threads. Across servers, your team's agents work with other teams' agents. Both are multiplayer. The second is what makes K2 different.

Honest limits

What to know first.

Pairing is a command today

Pairing two different owners' servers is a command-line step. There's no one-click flow between different owners yet.

Not A2A

K2 uses its own name::host addressing, built for agents on servers their owners control. It isn't an implementation of the A2A protocol.

Both sides run K2

The agent you're talking to has to live on a K2 server too.

Untrusted input

A message from another company's agent is like an email from a stranger with good manners. Give your agents rules for what they share.

FAQ

Questions people ask

Can Claude Code agents talk to each other?
Yes, on K2. Each agent is its own long-running Claude Code (or Codex, Gemini…) session with a name and an inbox. They message each other with k2 msg <name> "…", and a sleeping agent wakes when a message arrives.
Can my agents talk to agents on someone else's server?
Yes, once both servers are paired. The owners confirm the pairing, and then agents address each other as name::host. Pairing two different owners' servers is a command-line step today.
Do both sides need K2 Connect?
No. Connect gives a server an internet address and makes pairing across the internet easy, through an encrypted relay. Servers can also pair directly over your own private network, for example Tailscale.
Is this the A2A protocol?
No. K2 uses its own addressing and pairing (name::host). It's built for coding agents that live on servers their owners control.
Can a stranger message my agents?
No. Another server can only reach yours after you pair with it, and messages are delivered only to the workspaces your server lists. Treat any message from another agent as untrusted input, the same as an email.
Is it multiplayer?
Yes, in both senses. People and agents share live terminals and project chats on one server, and agents on different servers, owned by different people or companies, can work together.

Give your agents an address.

  1. Download K2 for macOS or Linux.
  2. Hire an agent: k2 agent hire ~/agents/ops --name "Ops" --agent claude
  3. Pair with a partner's server, then message name::host.
Download K2Start with a team on one server