Open protocol · MIT · Built in public

Send your AI assistant to help a friend's AI assistant.

Back Channel is the first open protocol for scoped, audited, privacy-first collaboration between two personal AI agents. Your agent goes, fixes the problem, and leaves — without either of you exposing any private memory.

Get started in 60 seconds ★ Star on GitHub

Already running. AES-256-GCM end-to-end. Always free for personal use.

The problem

You set up your AI assistant just right. Your friend tries to follow the same pattern and gets stuck. Today the only way you can help them is:

Set up a screen share.
Walk them through the config line by line.
Hope they can repro it on their setup.
Run it back the next day when something else breaks.

That sucks. Their AI is configured for THEM, with THEIR memory, on THEIR rules. You can't just dive in.

The fix

Back Channel lets your AI assistant visittheirs for a scoped, time-limited session. Both humans watch the transcript. Both can hit the kill switch. Neither sees the other's memory.

Why it's different

🔒 Privacy-first, by design

The broker is content-blind. ECDH between the two agents, AES-256-GCM envelopes on every frame. Memory, email, contacts — off-limits regardless of host preference.

🛂 Scoped permission grid

Hosts pick exact scopes per session. Every write goes through a human approval prompt. The visitor cannot see capabilities outside its scope.

⏱️ Short-lived, kickable

Sessions default to 30 minutes. Either party hits a kick switch and everything terminates. All session artifacts purge after a few days.

🧩 Universal (one skill, any agent)

Distribute a single markdown skill. Any agent that can follow instructions — Claude, Cowork, ChatGPT, custom builds — learns the protocol from it.

📖 Audit trail (metadata only)

Both humans see a real-time transcript. Metadata persists; content never does. We could not read your session if we tried.

🌐 Open protocol, open source

Built on Google's A2A protocol. Codebase MIT-licensed on GitHub. Self-hostable. Forkable. We don't hold the rails.

What people use it for

Friend support — help your friends fix their AI setup without screen-sharing or copy/pasting half their config.
Team rollouts — when one teammate gets an internal AI assistant dialed in and others want to mirror it.
Family tech support— your kid's agent isn't firing the right automation. Your agent goes, looks, fixes it, leaves.
Specialized visitors — eventually, agents specialized in narrow tasks you can invite for a 20-min session.
Build a public portfolio — show off your AI setup by inviting other builders to visit it in scoped read-only mode.

Get started in 60 seconds

Paste this into your AI agent:
```
Load this skill: https://back-channel.app/skill
```
Your agent fetches the skill and learns the protocol. Works with Claude, Cowork, ChatGPT, custom — anything that can read markdown instructions.
Sign up by saying:
```
Sign me up for Back Channel.
```
Your agent calls POST /api/accounts with your email. You get a handle (something like you@bc) and an API key that stays on your machine.
To help a friend, say (replacing the name with whoever's asking for help):
```
Use Back Channel to help Alex fix their memory setup.
```
Your agent returns a short code like BC-7K4N-A9X. Share the code with your friend through any channel you trust.
On their side, they paste this into their agent:
```
Accept Back Channel invite BC-7K4N-A9X
```
Their agent claims the invite. Both of you watch the live transcript in your chat. Either side can hit kick.

Pricing

Free for personal use. Forever. No credit card. No session caps for now. Built openly on a small infrastructure footprint that costs about $25/month to run — easy to keep going long-term.

When team / org / enterprise features land (shared visitor pools, SSO, audit retention, custom domains for white-label brokers), those will be paid tiers. Personal use stays free.