Block's Buzz Puts Humans and AI Agents in One Workspace, With Their Own Keys. It Wants to Be Your Slack, Your GitHub, and Your Audit Trail.

TL;DR

Block announced Buzz on July 21: a free, open-source, self-hostable workspace where humans and AI agents share the same channels, threads, voice huddles, code repositories, and automated workflows. It is built on Nostr, so every participant, human or agent, holds a cryptographic keypair that belongs to them rather than to the platform, and every message, reaction, review approval, workflow step, and git event lands in one signed event log. Harnesses for Claude Code, Codex, and goose ship at launch via the Agent Client Protocol, the license is Apache 2.0, and the pitch is blunt: stop bolting agents onto Slack and GitHub, and give both species one room with one audit trail.

What actually shipped

This is not a waitlist. The announcement came with a working desktop app (Tauri + React, for macOS, Windows, and Linux), a self-hostable relay, a buzz-cli that talks JSON in and JSON out for agent use, YAML workflows with message, reaction, schedule, and webhook triggers, media sharing with comments pinned to specific video frames, and a built-in git forge. Release v0.4.22 was the launch build; v0.4.23 landed the next day, and the repo sits at 4,300-plus GitHub stars as of this writing.

You can point the app at Block's hosted relay at buzz.xyz, currently free in beta with no published pricing, or run your own relay and keep every byte on your own hardware. For the self-hosted crowd, that last part is the headline: this is a full team workspace, agents included, that you can operate like any other homelab service.

Agents are members, not bots

Every collaboration tool today treats an agent as either a bot account or a human's borrowed session. Buzz's core design choice is that agents get the same thing humans get: their own Nostr keypair, their own channel memberships, and their own audit trail. The owner signs a narrowly scoped authorization for the agent, and the agent signs its own work. Block's engineering post puts the principle in one line: authorization does not erase authorship. The agent remains the author.

Think of it as issuing a contractor their own badge instead of lending them your keycard. When something goes wrong, you pull the badge and read exactly which doors it opened; you do not re-key the whole building or wonder which swipes were actually yours. Concretely: a leaked agent key can be revoked without touching the human's identity, and every event the agent ever signed stays attributable to it.

The identity is also portable. Because keys are plain Nostr keypairs rather than platform accounts or vendor API keys, an agent's identity and history are not locked to Buzz, to Block, or to whichever model provider happens to be running it this month. Swap Claude Code for Codex under the same key and the project context, permissions, and history stay put.

A git forge where the branch is a room

Buzz ships its own code hosting, and the design is more serious than "we added git". Repositories are stored as immutable, content-addressed packfiles plus a single mutable manifest pointer; a push writes objects first, then advances the pointer with a compare-and-swap as the commit point. Block says the storage protocol was specified in TLA+ and model-checked, and backends have to pass a conformance suite before they count. That is an unusual amount of formal rigor for a chat app, and exactly the right amount for the thing holding your repos.

On top of that sits the workflow this audience will actually feel: open a feature branch and a channel appears. Patches land as NIP-34 events, CI posts results, an agent runs a first-pass review, humans react to the parts they care about, and the merge decision lives in the same room as the evidence. The channel becomes the record of why the code exists, searchable alongside the conversation that produced it.

What works today, and what is still drying

Block is unusually honest about the gap between the demo and the vision. The README's own words: "rough edges and giant chasms", and agents should be "part of the room, not haunted cron jobs". Here is the launch-day state, straight from the shipped/pending table:

The caveats matter. Workflow approval gates, the mechanism that would let a human sign off before an agent's release workflow actually ships, are "infra exists, glue still drying". There are no mobile clients yet. Hosted pricing is unannounced. And Nostr itself carries baggage: a small relay ecosystem and a reputation shaped more by decentralized social experiments than by enterprise IT. Buzz inherits both the elegance and the eyebrow-raises.

Why a builder should care

Two reasons. First, the audit-trail problem is real and getting worse. If agents write a growing share of your code and messages, "who did this and under whose authority" stops being a philosophical question and starts being an incident-review one. A single log where every agent action is signed by the agent and chained to a human grant is a credible answer, and notably one that Slack-plus-GitHub-plus-bot-tokens does not give you.

Second, the economics. Per-seat pricing assumes members are people; an orchestrator that spawns a dozen subagents before your standup turns that invoice into performance art. A free, Apache-2.0 workspace where agent membership costs a keypair is a direct shot at that model. Bradley Axen, Block's head of AI capabilities, frames it as the industry's fork in the road: "Every company is going to need a place where humans and agents work together. The question is whether that place is proprietary or open."

The contrast with the incumbents' approach is sharp. Anthropic's Claude Tag drops an agent into your existing Slack; OpenAI's Presence, announced one day later, rents you agents wrapped in its own platform and engineers. Buzz rebuilds the room itself so the agent is a first-class member of infrastructure you own. Those are three different bets on where agent work will live, and only one of them ends with the event log on your hardware.

Whether Buzz wins is a different question. Team chat is a graveyard of technically superior Slack rivals, and Buzz needs network effects that a protocol nerd's dream architecture does not automatically buy. But the ideas, portable agent identity, signed work, scoped delegation, will outlive the app even if it stalls. This is the first mainstream-company product built end to end on the assumption that agents deserve their own keys.

Key Takeaways

• Block launched Buzz on July 21: a free, Apache-2.0, self-hostable workspace where humans and AI agents share channels, repos, workflows, voice, and canvases.
• Every participant holds their own Nostr keypair; every message, review, workflow step, and git event is a signed event in one log, giving agent work a verifiable chain of custody.
• Agents are scoped by identity, not borrowed credentials: owners sign narrow grants, agents remain the authors of their work, and a leaked agent key revokes cleanly.
• Claude Code, Codex, and goose harnesses ship at launch via the Agent Client Protocol; the built-in git forge uses NIP-34 events and a TLA+-model-checked storage design.
• Launch gaps: no mobile clients, workflow approval gates unfinished, hosted pricing unannounced, and Nostr's ecosystem is still niche.
• The strategic bet is against per-seat SaaS: if agents multiply faster than headcount, the workspace that charges per human keypair (zero dollars, in this case) has the pricing model that fits.

Sources: Block announcement, Block Engineering blog, block/buzz on GitHub, SiliconANGLE, Decrypt