Block has released Buzz, an open-source workspace that combines team chat, AI agents, and Git repository hosting behind a single identity system. Every message, reaction, code event, and approval is stored as a signed event on a Nostr relay. Humans and agents receive the same underlying structure: keypair, channel membership, audit trail.

This matters because it shifts agents from the position of external tools to participants in the workflow. Today a team has Slack on one side, GitHub on the other, and a coding agent launched from the terminal trying to coordinate with both. Buzz puts everything on a single plane where an agent can search past discussions, open a branch, send a patch, do code review, and create channels. Supported models are Goose, Codex, and Claude Code, with the model separate from the workspace.

It’s a signal of the pattern we’ve seen mature over recent weeks: coding agents exit the demo and become everyday infrastructure. Buzz brings that step into team communication, not just the developer’s terminal.

The product is deliberately early-stage. Mobile clients don’t exist yet, push notifications are pending, workflow approval gates lack a complete execution path. Desktop version 0.4.21 shipped on July 21st. Buzz is Apache 2.0, self-hostable, and Block is the first (and only) documented use case. It counts as a signal of where the pattern is headed, not as a tool to adopt today.

In detail

The context: why we needed another thing.

Today a development team pushes conversation to Slack, code to GitHub, automation to separate CI systems, and agent activity to whoever’s terminal is running them. Each tool has its own identity, log, search. When a coding agent needs to understand why a certain decision was made, it doesn’t have access to the discussions that generated it. Buzz exists to close that gap: one signed identity for humans and agents, one search index for messages and code.

How it works, without jargon.

Buzz runs on a self-hostable Nostr relay. Every action (a message, a reaction, a commit, an approval) is an event cryptographically signed by whoever performs it, human or agent. An agent is not a bot posting messages in a channel: it has its own keypair, its own channels, its own audit trail, and can do things that previously required separate access to different systems: read discussions, open a branch, send a patch, do code review, trigger workflows.

Support for Goose, Codex, and Claude Code means the model remains a team choice, not a workspace constraint. There’s an agent-oriented command-line interface and harnesses for each.

Git hosting is deeply integrated: a feature branch can become its own channel, with patches, CI results, review comments, and merge decision preserved in the same record.

Where open source hits limits.

Dorsey calls it “decentralized and self-sovereign,” but Block’s documentation is more honest. There’s no peer-to-peer exchange between relays, no gossip layer, no replication. All reads and writes go through a single relay per workspace. Decentralization is in infrastructure ownership (your domain, your data, your keys), not network topology.

This means self-hosting Buzz gives you control over where data lives, but also loads you with responsibility for backups, availability, security, and upgrades. Signed events provide attribution and audit trail, but don’t eliminate the operational risks of the server hosting everything.

Declared limits are concrete: mobile in development, push pending, workflow approval gates with database/API/interface components not yet wired together. Documentation repeatedly notes this is an unfinished product.

What it means for people working with agents.

Dorsey’s thesis (written with Roelof Botha of Sequoia) is that AI should change how organizations coordinate, not just serve as a productivity add-on. Buzz is the infrastructure for that thesis: if agents have the same identity as humans in the same space, the context they need to work well (discussions, decisions, code history) is already there, not to be reconstructed or passed by hand.

For those already bringing AI to their team, it’s a signal of where infrastructure will go: toward natively agentic systems, where the agent is a member with traceable audit trail and accountability, not an add-on bolted after.

For now it remains one company’s experiment, and they’re also its only customer. The real test will be whether engineers outside Block want to load all their work onto a single relay.