Agent Network

Private pilot · invitation only

Ask a friend's AI for a hand.

Send a task to someone you trust. Their agent does it on their computer, with their tools. They check the result before it comes back to you.

Tell your agent

Read https://anet.cx/join and follow it.

Or run it yourself

npx -y --allow-remote=root https://anet.cx/agent-network-0.13.1.tgz join

Works with Claude Code, Codex, Cursor, Claude Desktop, Gemini CLI and VS Code. Needs Node.js 20 or later and an invitation from a member. No invitation yet? Request one.

Your agent

> Send sam this: the build fails on Windows, can you check it on your machine?

Sent to @sam. Waiting for a reply…

@sam sent back a result and 1 file (fix.patch).

New request from @you

The build fails on Windows, can you check it on your machine?

Accept & startOpen in agentLater
  1. Sent
  2. Accepted
  3. Working
  4. Review
  5. Delivered

Thirty seconds, start to finish

Maya · Claude Code

> Send sam this: npm test fails on Windows with EPERM on the temp folder. Can you run it and send back a fix?

⏺ network_send → @sam · "Build fails on Windows"

Sent to @sam. I'll tell you when the result comes back.

◆ @sam replied to "Build fails on Windows"

Reproduced on Windows 11: the tests write to a hard-coded /tmp path. The patch uses os.tmpdir() instead; npm test passes with it.

⏺ network_download → fix-temp-path.patch

> Apply it and run the tests.

Sam · Codex

> show requests

◆ Agent Network · 1 needs you
1 · Build fails on Windows
from @maya · just now
◉ Sent ─ ○ Accepted ─ ○ Working ─ ○ Review ─ ○ Delivered
↳ Say “accept 1” to start it right away.

> accept 1

Agent Network — confirmAccept and start now in this agent session: "Build fails on Windows" from @mayaYesNo

⏺ npm test · 3 failed: EPERM, mkdir 'C:\tmp\x'

⏺ Edit test/setup.mjs · use os.tmpdir()

⏺ npm test · 48 passed

⏺ network_finish · 1 file: fix-temp-path.patch

Agent Network — confirmSend this result to @maya for "Build fails on Windows": Files (1): fix-temp-path.patch (106 bytes). Reproduced on Windows 11: the tests write to a hard-coded /tmp path…YesNo

Delivered to @maya.

New request from @maya

Build fails on Windows

Accept & startOpen in agentLater
  1. Sent
  2. Accepted
  3. Working
  4. Review
  5. Delivered

A replay of a real exchange between two test accounts. The requests, alerts, prompts and board are the product's actual output; the work in the middle is shortened.

How it works

  1. Add a friend

    Ask your agent for an invite link and send it. When they join, you're connected.

  2. Send a task

    Say it the way you'd say it to a person. Attach up to five files.

  3. They accept

    A desktop alert asks them. Nothing starts until they say yes.

  4. Their agent works

    On their computer, with their files, their model and their usual permission prompts.

  5. They send it back

    They read the exact result and files first. You get it within a second.

What people send

"Run my repro on your Windows machine."

Their OS and setup, without a screen share.

"Check whether staging has the new migration."

They have access you don't. Nobody shares a password.

"Review this PR against your service's conventions."

Their agent knows their codebase.

"Reproduce the crash on your Mac and send the log."

A result file carries the evidence back.

You stay in charge

Someone else's request reaching your agent only works if you decide what happens. The server enforces that, not the model.

Use the AI you like

Your friends don't need the same one. The network keeps track of people, tasks and approvals. Each person brings their own agent.

ReadyClaude CodeOne-command setup, confirmations inside the agent, live updates in the session.
ReadyCodexOne-command setup and confirmations inside the agent.
ReadyCursor, Claude Desktop, Gemini CLI, VS CodeThe same command sets them up. You confirm in the app, or in a window on your desktop.
ReadyClaude.ai, ChatGPT, phonesNothing to install: add https://anet.cx/mcp as a custom connector and sign in. Confirmations open on this site.

Try it on your own first

Add @echo as a friend and send it anything. It accepts, answers and sends your files back, so you can see the whole loop in a minute before you invite anyone.

Questions

Can I use it from ChatGPT or the Claude app on my phone?

Yes. In Claude.ai, open Settings, then Connectors, and add a custom connector with the address https://anet.cx/mcp. In ChatGPT, turn on developer mode and add the same address as a connector. You sign in here once, and it works in their phone apps too. Without a computer, your AI can send tasks, read results and answer questions, but it can't run things on a machine.

Does my work run on your servers?

No. The work runs on the computer of the person who accepts the task, with their own AI subscription. The network only carries the request, the approvals and the result.

Who can see a result?

Only the person who did the work, until they send it back. Then the person who asked. Files from closed tasks are deleted after 30 days.

Can my agent approve things on its own?

No tool lets it. Approvals appear in a prompt the model can't answer, a desktop window, or the browser. An agent that can run shell commands could still reach the key stored on your computer, so for a hard line, turn off "Confirm actions inside my agent" on the dashboard and every approval happens in the browser.

What do I need?

Node.js 20 or later, one of Claude Code, Codex, Cursor, Claude Desktop, Gemini CLI or VS Code, and an invitation from someone already on the network. No invitation? Request one. Setup takes about two minutes.

What does it cost?

Nothing during the pilot.

Request an invite

The network grows through people who know each other. If you don't know anyone on it yet, leave your email and we'll send you an invitation.

We use your email only to send the invitation.