Verdict: Buzz is the most serious open-source attempt yet to give AI agents a workspace of their own — not a pinned bot inside a chat tool, but a self-hostable collaboration platform where every human and every agent carries its own cryptographic identity. For a small business or solo builder already running Claude Code or Codex, Buzz is free to try at buzz.xyz today and genuinely useful for orchestrating multi-agent workflows — but it shipped on July 21, 2026 in an early beta that Block's own team describes with the words "rough edges and giant chasms." It's worth a weekend evaluation; it is not yet where you should migrate your whole company.
Last verified: 2026-07-30 · Free everywhere (hosted + self-host + source). Apache 2.0. GitHub stars: ~16.9K. Late beta ("rough edges"). Not production-grade.
What is Buzz, and who built it?
Buzz is a free, open-source collaboration platform released on July 21, 2026 by Block — the company led by Jack Dorsey (co-founder of Twitter and Square, now Block). Its public site is buzz.xyz, and the full source code lives at github.com/block/buzz under an Apache-2.0 license. Dorsey introduced it in a post on X as "a new groupchat platform for teams of people and agents of all sizes… model-agnostic, decentralized, self-sovereign, and open source." [1]
The core idea is simple: what happens when an AI agent becomes a full member of a workspace instead of a side panel you paste into? Block spent two years building internal agent tooling and landed on this thesis — paraphrased from Bradley Axen, Block's Head of AI Capabilities: "Every company is going to need a place where humans and agents work together. The question is whether that place is proprietary or open. We built Buzz because we believe it should be open." [1]
The interface will feel familiar to anyone who has used Slack or Discord: channels, threads, direct messages, voice huddles, media sharing, search, and code repositories. The difference is who is in the room. Agents hold their own portable cryptographic identities (built on the Nostr protocol), they have scoped and revocable permissions, and every message, patch, and workflow run they make lands in one signed, auditable log. [1][2]
Is Buzz actually free? What's the catch?
Yes — as of July 30, 2026, Buzz is free everywhere it exists, and Block has published no paid tier or enterprise pricing yet. [1][3]
| Deployment | Cost | Limits | What you bring |
|---|---|---|---|
| Block-hosted (buzz.xyz) | Free | Invite-only communities, 5 GB media storage, 365-day content retention, max 3 communities per account [3] | Just an account |
| Desktop app (macOS .dmg, Linux .AppImage/.deb, Windows .exe) | Free | Point it at a relay via BUZZ_RELAY_URL |
A relay to connect to (yours, Block's, or a community's) |
| Self-hosted from source | Free (your infrastructure) | Whatever your hardware supports | Docker + Rust 1.88+ + Node 24+ + pnpm 10+ + PostgreSQL + Redis + S3/MinIO [3][4] |
The catch is that Block treats the hosted version explicitly as an "early stages" invitation (the site literally says "come test the early stages with us"). There are no published SLAs, no announced usage caps roadmap, and Block itself describes the current state as "useful today" but with "rough edges and giant chasms." The first independent week-long field report (DevToolsDaily's deployment, July 26) confirmed that agents do do real work but identified a serious observability gap: once an agent receives a message, the UI shows you nothing about what happens next. [3] Treat "free" as a launch-phase fact, not a permanent promise.
How does Buzz compare to Slack plus GitHub plus a coding agent?
Buzz's real pitch is that it collapses three separate systems — chat (Slack), a code forge (GitHub), and your agent harness's terminal context — into one signed event log where a human and an agent hold the same kind of identity.
| Dimension | Slack + GitHub + Claude Code/Codex (status quo) | Buzz |
|---|---|---|
| Agent identity | A bot token, impersonating a user or sewn into a single runtime's account | Own portable cryptographic keypair, scoped + revocable, platform-independent |
| Agent context | Lives in your terminal or inside one harness's context window; your team's discussion about that work lives in Slack — two separate worlds | Same log: the conversation, the patch, the workflow run, and the approval are one kind of signed event |
| Audit trail | Split across Slack logs, GitHub webhooks, and your CLI history | One append-only, hash-chain-backed log; every action cryptographically signed |
| License + self-hosting | Proprietary (Slack) + historically closed-ish (GitHub); you cannot run Slack on your own computers without paying | Apache-2.0; run it on your own relay and nothing touches anyone else's servers |
| Maturity | Mature, fast, battle-tested | Launched July 21, 2026; "rough edges and giant chasms" per Block itself [1] |
| Model choice | Slack doesn't care (it just sees a bot), but your agent harness's model is what matters | Model-agnostic and harness-agnostic — bring Claude Code, Codex, Goose, or build your own [1][3] |
The decision frame is not "Buzz vs Slack for chat" — on that axis, Slack is far more mature. It's "do I want agents to be first-class signed participants in one event log, or remain second-class symptoms bolted onto three different platforms?" The model-agnostic stance is the thing that makes Buzz interesting for builders: the workspace doesn't care whether your engine is Claude, Codex, xAI, Gemini, or Block's own goose (an open-source Apache-2.0 agent from Block consumers' AI Foundation).
How do you actually set up an AI agent team in Buzz?
The practical flow, once you have an account or have connected the desktop app to a relay:
- Create or join a Buzz community. A community is the top-level container — one relay URL, one identity model. Hosted communities on buzz.xyz are invite-only and include 5 GB of media storage with 365-day retention. [3]
- Set up a channel per workflow. Think of each channel as an always-on office for the work — one for your image generation pipeline, one for blog drafting, one for SEO keyword research, one for code review. Channels can be private or open; threads live inside them.
- Create or import an agent. Each agent gets its own profile, instructions, and a model. You can import a team snapshot or build one from scratch. The
buzz-acpharness (ACP ↔ MCP bridge) lets you plug Goose, Codex, or Claude Code agents directly in. [3] - Give the agent its tools and skills. If you've already taught Claude Code how to post to WordPress or work with your Obsidian vault, those skills carry into Buzz — the harness inherits what you've trained upstream. This is a real productivity unlock for builders already invested in a coding-agent setup. (For a closer look at how that style of memory/instruction inheritance works, see our guide to building a free AI agent operating system with Hermes Agent and Obsidian.)
- Tag the agent into a channel and give it a task. The interaction reads like pinging a teammate on Slack —
@codex generate thumbnails for the latest video,@claude draft the SEO outline for next week's post. The agent reads its own context, proposes work, asks for the missing pieces, and acts. You see the result and the code in the same flow. - Add humans where a human really should approve. Humans join channels as regular members. For workflow steps that need sign-off — "ship this to production", "publish this post", "merge that PR" — you keep the human in the loop alongside the agent, rather than trusting the agent to fire off writes blindly. (If you're already running a personal agent fleet, our Hermes Agent fleet setup walkthrough covers the same pattern of mixing human and autonomous steps.)
- Review the audit log. Every action by every agent is a signed event in one hash-chain-backed log. The search is full-text across messages, patches, and workflow runs — you ask the workspace "what did codex change last week?" and you get the actual threads, not vibes. [3]
What practical workflows does Buzz unlock for small teams?
The honest answer is: anything you currently keep in a flaky checklist, a shell script, or a separate codehawk automation session. Buzz turns these into persistent channels that remember their own history and have agents sitting in them ready to go.
Content + publishing. One popular pattern (one that works even if you publish to anything — WordPress, Substack, or your own Next.js blog) is a channel where an SEO agent drafts the post outline, a writing agent generates the full draft, and the human reviews before publish. The agent inherits your existing style and formatting skills and posts the result directly to your CMS without you logging in. We've previously documented how to run OpenAI Codex for free across multiple providers — Buzz is the same "use the engine you already pay for" philosophy, but extended to the whole workspace layer rather than just the CLI.
Image and thumbnail generation. You feed examples into the channel, tag the agent, and ask for variants. The agent ingests the reference set, diagnoses what's wrong with the current image ("too low-contrast", "wrong text hierarchy"), offers several directions, and generates new versions — all without you leaving the chat.
Code review + repository work. Buzz ships an early Git forge built on NIP-34 (the Nostr patch / repo spec). Repositories are stored as content-addressed packfiles on object storage. Block publicly says the forge's storage design was specified in TLA+ and model-checked — a serious engineering signal, even if the UI is early. The long-term vision is an open-source community hosting its own code and its conversations and its agents in one place. [1][3]
Cross-harness orchestration. Because Buzz is model-agnostic, you can have a cheap model do triage and reference generation and a more expensive model make the real decision — within the same workspace, without manual hand-offs between terminals.
What's still broken or missing?
Many things, candidly — and the value of this section is that it's drawn from Block's own acknowledgments plus the first week of public field reports.
- No agent observability. Once an agent receives a message, you don't see what it did between that and its reply. This is the single biggest operational gap. [3]
- Mobile clients in flux. An Android app is on Google Play (
xyz.block.buzz.mobile); there is no iOS App Store listing as of July 25, 2026. Push notifications are "pending". [3][4] - Workflow approval gates are descriptions, not executions. The docs describe the gate as having database, API, and interface components without a finished execution path. In practice the fact that approval gates don't really work yet means you can't safely let agents ship the last mile without a human in the loop, which is the opposite of the autonomy the project promises. [3]
- Centralized relays in practice. Dorsey's launch called Buzz "decentralized, self-sovereign, and open," but there is currently no peer-to-peer event exchange, gossip layer, or replication between relays — every read and write in a workspace passes through one relay. "Self-sovereign" today means closer to self-hosting with portable identity than a fully decentralized event mesh. The Hacker News launch thread (200+ comments) pushed back hard on this framing. [4]
- No published SLAs, customer adoption numbers, or a pricing roadmap. Block acknowledges Buzz by being its first user and its first documentation, and there are no published third-party enterprise customers yet. [4]
- 796 open issues on GitHub as of late July — ~708 commits in the previous four weeks is a sign of healthy momentum, but also a sign of how rough a project this young and this active is. [4]
What this means for you (small-team / solo-builder read)
If you're a small-business owner, freelance builder, or content creator already using AI agents for any part of your work, here's the honest decision:
- Do evaluate Buzz this month. It's free, sign-up takes five minutes, and the model-agnostic, self-hostable premise is the right architectural bet for where agents are going in the next 18 months. Try the hosted version at buzz.xyz with one workflow you currently run by hand — say, blog drafting or image generation — and see whether agent-as-teammate fits how your team actually works.
- Don't migrate your whole company yet. Slack is more mature; GitHub is more mature; your agent harness's own terminal already works. Buzz is the right bet, not the right migration, in July 2026.
- Pick an angle that fits your stack. Open-source users and self-hosters benefit most from Buzz today because you can run it on your own relay with your own Agent Communication Protocol adapters — no third-party servers, no vendor lock-in. Hosted users benefit from Buzz as a free Slack replacement with agent teammates, but the first independent week-long report flags the observability gap as the pain point.
- Keep your existing coding-agent subscription. Buzz doesn't charge you for model usage — it just lets the agents you've already paid for (Claude Code, Codex, Goose) collaborate through one shared workspace. That is the real "free" angle: the workspace itself costs nothing; you bring the models.
- Watch the commitment cadence. The GitHub repo is shipping at 681+ commits per four-week block with releases like v0.4.26 that ship real stability fixes. The project will look different by September 2026 — re-evaluate on a 60-day cadence. [4]
Buzz compared to the next layer of your agent stack
Where does Buzz sit among the dozen AI-agent surfaces you might already use? The adjacent layers:
| Surface | What it does already | Overlap with Buzz | How they fit together |
|---|---|---|---|
| Claude Code / Codex / Cursor (coding agent) | The agent itself — terminal, editor, or cloud; does the work | Agents are the engines that sit in Buzz as members | Buzz wraps your engine's context into a shared, auditable workspace; your engine still does the coding |
| Slack / Discord (chat) | Team communication | Same channel / thread / DM surface vocabulary | None fundamental — Buzz explicitly tries to be this layer |
| GitHub / GitLab (code forge) | Code hosting + PR workflows | Buzz has an early (TLA+-modeled) Git forge built on NIP-34 | Buzz's forge is pre-1.0; don't retire GitHub this year |
| Hermes Agent / Personal agent runtime | Long-running autonomous loops with tool calling, memory, scheduling | Hermes-style runtimes are the engine; Buzz is the surface | Plug a Hermes-agent runtime as one of your agents in Buzz (via the ACP / MCP bridge); see our walkthrough of building a free agent OS with Hermes + Obsidian for the engine layer Buzz can sit on top of |
| Free open-source AI tools on GitHub | The wider catalog | Buzz is itself one of these — Block's Apache-2.0 entry | Buzz is a modern addition to that same long-running open-source movement |
| Your productivity setup (6-step agent workflow) | How you organize your personal agent stack | How you organize the work Buzz helps you share | Use Buzz's "one channel per workflow" structure to mirror your personal setup at a team scale |
Poweruser tip: the single most useful shared-documentation benefit Buzz offers — once you've wired everything in — is that your agent's internal memory/instructions can be carried in the agent profile sitting in the channel, not buried in a context window nobody else can see. The full productivity orientation for that kind of agent-team thinking is in our guide to setting up AI agents for productivity (2026).
FAQ
Q: What is Buzz by Block? A: Buzz is a free, open-source (Apache 2.0) team workspace launched by Block on July 21, 2026, where humans and AI agents share the same channels. It looks like Slack — channels, threads, DMs, voice — but agents join as members with their own cryptographic identities, and every message, code patch, and workflow step is a signed event in one auditable log, built on the Nostr protocol. [1]
Q: Is Buzz free, or is there a paid plan? A: Everything is free as of July 30, 2026. The code is Apache-2.0, Block's hosted version at buzz.xyz is free in early beta, and self-hosting is free on your own infrastructure. No usage limits, paid tiers, or enterprise pricing have been published yet — treat "free" as a launch-phase fact, not a permanent promise. [1][3]
Q: Which AI agents and models can I use with Buzz?
A: Buzz is model-agnostic and harness-agnostic. You can plug in agents built on Claude Code, OpenAI Codex, or Block's own Goose (Apache-2.0 open-source) via the buzz-acp bridge (ACP ↔ MCP). You can also bring your own agent or build a new one inside Buzz. The workspace doesn't prescribe what's behind the agent — it just provides the place where agents and humans collaborate. [1][3]
Q: Is Buzz a blockchain or crypto product? A: No. Buzz is built on Nostr, a lightweight open protocol for cryptographically signed, relayed messages — the same protocol Dorsey has backed for years. Nostr is not a blockchain; there is no token, no consensus, no chain. Buzz uses Nostr for portable cryptographic identity and signed event storage. [1]
Q: Can I self-host Buzz on my own infrastructure?
A: Yes. Clone github.com/block/buzz, run just setup && just build && just dev, and the relay starts on ws://localhost:3000. Self-hosting needs Docker, Rust 1.88+, Node 24+, pnpm 10+ (or Block's bundled Hermit toolchain), PostgreSQL, Redis, and S3/MinIO. Nothing routes through Block when you self-host. [3][4]
Q: Is Buzz ready for production or team migration? A: Not yet. Block itself calls the beta "useful today" with "rough edges and giant chasms," and the first independent week-long field report confirmed agents do real work but flagged a serious observability gap — you can't see what an agent did between receiving a message and its reply. Workflow approval gates and push notifications are incomplete. Evaluate Buzz as a bet, but don't retire Slack or GitHub yet. [1][3]
Q: What's Nostr and why does it matter here? A: Nostr ("Notes and Other Stuff Transmitted by Relays") is an open protocol where every event (message, patch, workflow run) is cryptographically signed by its author's keypair and relayed to clients. In Buzz this means every participant — human or agent — holds a portable identity that doesn't belong to any platform. Your agent's identity and history travel with the holder, not with Block. [1]
Discussion
0 comments