You add an AI agent to a channel the same way you add a person. It posts, and it signs the same kind of event a human signs. That small detail is the whole idea behind Buzz, the Rust chat app that Block, Jack Dorsey's company behind Square and Cash App, recently open-sourced.
Most agents live in your editor, blind to the review, the argument, and the call to ship. Buzz puts them in the same rooms as the people.
I've spent real time in it, and the honest version is more interesting than the launch pitch.
What it is
Picture Slack with a git forge attached, self-hosted. That's Buzz. The twist is that the AI agents in it are members, not bots. Add one and it can open repos, send patches, review code, and run workflows.
It's written in Rust, Apache licensed, and built on Nostr, the signed-event protocol Dorsey has funded for years. As I write this, the latest release is 0.4.22, so it's early, and it shows.
How it works
The same idea repeats everywhere, because every message, reaction, review approval, and git event is a signed event in a single log. That way, a chat message, a code patch, and an approval are all the same type of record, so one search can find them together.
Every one of those events is signed, and each agent signs with a key of its own. The docs call that key a passport instead of an API key. The owner grants what the agent is allowed to do, and if a key leaks you revoke that one agent without touching anyone else.
Underneath it all sits one relay, backed by Postgres for events and search, Redis for presence, and object storage for media. There's no hosted demo, so you host it yourself. A source build wants Docker and the pinned Rust and Node toolchain, and the documented path is short.
git clone https://github.com/block/buzz.git && cd buzz
. ./bin/activate-hermit # pinned toolchain
just setup && just build
just dev # relay on ws://localhost:3000 + desktop app
Once the relay is running, your agents reach it through the buzz-acp harness, and Goose, Codex, and Claude Code all speak to it the same way. That harness is also the hard part at 0.4, and it still fails for people. On Windows the agent shell needs Git Bash, and the installer is unsigned.
What it is good for
That rough setup decides who it's for. The clearest fit is a small team that already self-hosts and wants agents doing real work, not just chat. The best example is branch-as-room, where a feature branch becomes a channel that holds the patches, the review, and the merge decision, so the room records why the code exists.
The second fit is a team that wants to own its identity instead of renting it from one vendor. Because every event is signed, Block argues, your history still verifies even if the company behind Buzz disappears.
The sharp edges
Buzz is pitched as decentralized, but today it isn't. Each workspace runs on a single relay that the code itself calls the single source of truth, with no peer-to-peer network and no replication behind it. So you can host it yourself, which is real, but hosting is not the same as decentralization. Nostr helps here with portable identity, yet it will not keep you online when someone wants to remove you.
Privacy is the harder problem. An agent that can read a channel reads all of it, and there is no clean answer yet for making sure it does not repeat what it saw to the wrong people. Its own key will tell you which agent leaked the data, but knowing that does not stop the leak itself.
There are a few more limits worth knowing. There is no end-to-end encryption, so the server can read all of it. The audit log is not tamper-proof either, because anyone who controls the database can rewrite it. And the workflow approval gates still do not survive a restart.
Verdict
Even with all those gaps, Buzz is genuinely good engineering. The weaker part is the Nostr and decentralization story around it, because it promises more than the tool delivers today. At first the keypair model looks like bot-per-user with a new name, and the maintainer has admitted as much. But the scoped, revocable keys and the shared audit trail are worth studying even if you never run the thing.
Adopt it now only if you self-host for fun and don't mind running a 0.4 release in production.
- Agents as members with their own keys is the idea to watch. The rest is packaging.
- Self-hostable, not decentralized. One relay per workspace, no replication.
- Multi-agent privacy is unsolved. Don't put a secrets channel next to a shared agent yet.
- No hosted demo, a heavy self-host, an unsigned installer. This one's for tinkerers today.
- Scoped, revocable agent identity will matter everywhere, even if you skip the app itself.