Block Open-Sources Buzz, Unifying AI Agent, Conversation, and Git Audit Logs Through Nostr-Signed Events

Block has released Buzz’s desktop client, relay, CLI, and agent harness, aiming to bring team chat, Git repositories, and agent workflows together in a single self-hostable workspace. Unlike approaches that attach bots to human accounts, each agent has its own key, channel membership, and activity history. Agents can open repositories, propose patches, review code, run workflows, and even coordinate other agents.

At the heart of Buzz is not another set of messaging tables, but a Nostr-style log of signed events. Messages, reactions, Git events, approvals, and workflow steps are all written to the relay using the same event model, with either a human or a program as the author. In theory, an audit system can use public keys to verify which agent initiated an action without relying on shared service accounts or chat logs maintained by model providers. The project comprises components written in Rust, TypeScript, and other languages, offers packages for macOS, Windows, and Linux, and supports running a self-hosted relay with Docker.

Agent integration is designed to be model- and harness-agnostic, supporting Claude Code, Codex, Block’s goose, and custom-built agents. The design’s real value for enterprise deployments lies in binding identity, permissions, and work products to the agents themselves. If different agents use different keys and channel scopes, revoking one agent does not require rotating human credentials at the same time, and it becomes easier to distinguish between model recommendations, tool execution, and human approval.

However, a signature proves only that a particular key submitted an event; it does not prove that the underlying model is safe or that its output is correct. The current single-relay architecture lacks cross-relay federation, replication, and failover. Messages and media on Block’s hosted relay are also not end-to-end encrypted, and external model providers may still receive channel context. Mobile clients, push notifications, and comprehensive approval gates are not yet mature. Engineering teams should initially treat Buzz as an inspectable prototype for agent collaboration, with testing focused on key custody, fine-grained authorization, data retention, and relay disaster recovery.