Verdict: For a small team already paying for Claude Code or a ChatGPT/Pro Claude subscription, setting Buzz up to share those coding agents across your whole team in shared channels takes under an hour with the desktop app and is the cheapest realistic path to multi-human-plus-agent collaboration today. (For a separate take on whether Buzz is overall production-ready, see our earlier usability verdict on Buzz by Block.) Self-hosting the Buzz relay and wiring the buzz-acp harness for Claude Code or Codex is the harder, "own your data" route — it works, but it is early beta. Start with the hosted app, then decide whether self-hosting the relay is worth it for your team.
Last verified: 2026-07-30 · Free everywhere (hosted + self-host + source). Apache 2.0. Buzz's open-source repo had ~16.9K GitHub stars and ~1.6K forks as of July 30, 2026. Late beta — mobile clients, push notifications, and workflow approval gates are still incomplete. Pricing/limits change often — last checked July 30, 2026.
What is Buzz, in one paragraph?
Buzz is a free, open-source team workspace released on July 21, 2026 by Block (the company led by Jack Dorsey, co-founder of Twitter), where every human and every AI agent holds its own cryptographic identity and shares the same channels, threads, code repos, and workflows. It runs on the open Nostr protocol — every message, reaction, patch, or workflow step is a cryptographically signed event on a relay you own or that Block hosts for you — and the entire codebase is on GitHub at github.com/block/buzz under an Apache-2.0 license. [1][2]
Why use Buzz instead of just using Claude Code or Codex in your own terminal?
The single-tenant problem is real: your coding agent's context lives in your terminal, your team's discussion about that work lives in Slack, and the code itself lives in GitHub. Buzz collapses those into one signed event log where a human and an agent hold the same kind of identity — so your teammates can see what the agent is doing, tag it into a channel by topic, and keep one audit trail across the whole project. [3]
The other big win is cost. Today, if three people separately subscribe to Claude Code or a ChatGPT plan, your team pays three times. With Buzz, one harness install on one machine can serve the whole team's channels — your team members tag the agent in from their own devices using their own identities. The agent runs locally on whichever operator set up the harness. [3][5]
How much does Buzz cost vs Slack?
Buzz the software is free. The only thing you pay for is the LLM subscription that powers the agent you connect to it (Claude Code, Codex/ChatGPT, or Goose). Here is how that shakes out against Slack's published pricing.
| Platform | Free tier | Cheapest paid plan (per user/month, annual billing) | AI agent included? | Self-hostable? |
|---|---|---|---|---|
| Slack [4][5] | Yes — 90-day message history, 10 app integrations | Pro at $7.25/user/mo (annual) · Business+ at $15/user/mo · Enterprise+ custom | No — Slack AI is a separate $10/user/mo add-on [5] | No |
| Microsoft Teams [5] | Yes (basic) | Essentials at $4/user/mo; Office bundle at $12.50/user/mo | Copilot add-on, separate monthly fee | No |
| Mattermost [5] | Yes (self-host) | Professional ~$3.25/user/mo | No native agent harness | Yes |
| Buzz [1][2] | Yes — all features, no user cap | $0 software / bring your own LLM subscription | Yes — agents are full channel members | Yes |
| Goose via Buzz [1] | Yes — Apache 2.0 | $0 + whatever model/API calls cost (bring-your-own provider) | Yes | Yes |
The key insight: Slack's per-seat model punishes teams for giving every teammate access to AI. Buzz inverts that — the cost is your one LLM subscription, shared across the team, with no platform seat fee at all. (For a deeper framework on keeping your agent token bill low overall, see our Hermes agent free token-cost strategy.)
How do you set up Buzz with Claude Code or Codex?
There are two paths: the hosted desktop app (easy, ~15 minutes) and self-hosting the Buzz relay from source (hard, full control). The video's own lesson was blunt: getting it running from source via Claude Opus took over twenty minutes and was abandoned — use the app unless you specifically need to self-host. [L]
Option 1 — Desktop app (the fast path)
- Download the app from buzz.xyz. Builds are available for macOS (.dmg), Windows (.exe), and Linux (.AppImage / .deb). The latest tagged release at the time of writing was v0.4.21. [2]
- Create an account and a community. Buzz detects whether Claude Code and/or OpenAI Codex are installed on your machine and will offer to wire those up as your agent harnesses. If you only want to look around first, click Skip for now — you can attach a harness later. [6][L]
- Create a channel (for example, a channel named
content-ideasorbug-fix). Pick a channel type — Ongoing or Temporary — and public or private. [L] - Invite an agent into the channel. Buzz ships three starter agents — Fizz (the "maker," builds stuff), Honey (good at writing), and Bumble (good at researching). You can create a brand-new agent from scratch, give it a custom purpose, pick its harness (Claude Code, Codex, or Goose), choose its model, and set an optional profile picture. [L][3]
- @mention the agent in a message like *"@Fizz come up with some content ideas around open-source coding tools for today."' The agent replies in-channel, exactly like a human teammate would, and the whole conversation is signed and searchable. [L]
Each agent gets its own settings panel where you can set:
- Default harness — Claude Code, Codex, or Goose.
- Model — pick from the harness's supported models (e.g., Claude Opus 5, GPT-5.6) or specify a custom model name. [L]
- Who it can talk to — anyone, your team only, or an explicit allow-list. [L]
- Environment variables — set an API key, or any env var the agent's harness needs. [L]
- Concurrent conversations — whether the agent can hold multiple parallel conversations. [L]
Option 2 — Self-host the relay from source (the hard path)
Use this only if you need full data sovereignty. The prerequisites are Docker and Hermit (or, alternatively, Rust 1.88+, Node 24+, pnpm 10+, and just). [2]
git clone https://github.com/block/buzz.git && cd buzz
. ./bin/activate-hermit
just setup && just build
just setup copies .env.example to .env and runs just bootstrap automatically. To point a desktop client at your relay instead of Block's, set BUZZ_RELAY_URL before launching the app, or switch relays from inside the app's settings. [2]
How do the agent harnesses (Claude Code, Codex, Goose) actually work inside Buzz?
Buzz is harness-agnostic. Every agent runs on a harness you choose, the agent process runs on your own machine or server — not on Block's infrastructure — and Buzz is the room the agent works in. The component joining the two is buzz-acp, an ACP (Agent Communication Protocol) harness that listens for @mentions on the relay, prompts your agent, and posts the agent's reply back through the Buzz CLI. [3][7]
| Harness | Maintainer | Native ACP? | What it needs | Best for |
|---|---|---|---|---|
| Claude Code | Anthropic | Yes — wraps the Claude Agent SDK | A Claude subscription; set BUZZ_ACP_AGENT_COMMAND |
Teams already standardized on Claude Code skills, hooks, and subagents |
| Codex | OpenAI | Yes — Rust CLI, Apache-2.0 | OPENAI_API_KEY (NOT a ChatGPT subscription). Adapter logs a 426 Upgrade Required warning at startup — expected and non-fatal. Buzz gates codex-acp on a minimum version. |
Operators who prefer GPT-family models or an Apache-2.0-only CLI |
| Goose | Block → Linux Foundation (Agentic AI Foundation) | Yes — speaks ACP natively, no adapter needed | Lowest-friction default. buzz-acp with no BUZZ_ACP_AGENT_COMMAND set will spawn goose acp. |
Lowest-friction default and multi-provider setups |
The harness and model are per-agent settings, so two agents in one community can run different harnesses and models — one on Claude Opus, one on GPT-5.6 — side by side. This is the model-agnostic promise Dorsey announced on X. [3][1]
How do you build a "team of agents" in Buzz?
The practical pattern, drawn from the source walkthrough, works like this. (For the broader pattern of orchestrating multiple specialized AI agents on one screen — independent of Buzz — see our free multi-agent AI team orchestration stack guide.)
- Designate a specialty per agent. Create one agent for content/writing, one for research, one for building/executing (the three starter agents named Fizz, Bumble, and Honey map roughly to these). Or create brand-new ones with a written "purpose" field — for example, a "video-production" agent that can create images, draft scripts, and assemble clips. [L]
- Channelize your work. Make a channel per project or per workflow:
content-ideas,bug-triage,release-notes, etc. Tag the right agent into each. The agent's notification speech-balloon lights up when it's @mentioned. [L] - Use @mention, not replies, to summon agents. In the current build, replying to an agent without explicitly @tagging it does nothing — you must tag every time. This is a known UX wart and a real coordination cost when you have multiple agents. [L]
- Let agents publish to your stack. Agents inherit access from the local environment you set them up in. For example, an agent running under Claude Code can use your already-authenticated local credentials to publish drafts to your WordPress, post to a connected CMS via Netlify, etc. — provided the harness has those tokens in its environment. The source walkthrough showed a Fizz agent drafting a full SEO blog post in WordPress as a draft and offering to drop it into WordPress directly, because Claude Code already had the local WordPress login session. [L]
What still doesn't work in Buzz (as of July 30, 2026)?
Block's own GitHub README is unusually candid about what's not done yet. [2]
| ✅ Works today | 🚧 Still being wired up | 💭 Opinion, no code yet |
|---|---|---|
| Relay, channels, threads, DMs, canvases, media, search, audit log | Mobile clients (iOS + Android, being built in Flutter) | Web-of-trust reputation across relays |
| Desktop app (Tauri + React) | Workflow approval gates (infra exists, glue drying) | Push notifications |
buzz-cli (agent-first, JSON in / JSON out) + ACP harness (Goose, Codex, Claude Code) |
Huddle lifecycle events | Culture features |
| YAML workflows: message / reaction / schedule / webhook triggers | ||
| Git events (NIP-34: patches, repo announcements, status) | ||
| Git hosting backend |
A small team should plan around the green column and treat yellow as "nice if it lands." Do not depend on the red column for anything load-bearing.
How does Buzz's Nostr foundation actually help you?
Three things make a real difference for a team that cares about data control:
- Each agent holds its own cryptographic keypair that belongs to the agent, not the platform. Your agent's identity, history, and reputation are portable across any Nostr-compatible system. This means your "content-writer agent" doesn't vanish if Block turns off the relay — it can participate on any other Nostr relay. [1][2]
- Every action is a signed event in one auditable log. Chat messages, reactions, patch reviews, workflow runs — all the same kind of event, all the same identity model. You can search six months of project history — including what an agent did and why — in one place. [2]
- Branches become channels. Open a feature branch and Buzz auto-creates a channel; patches land as NIP-34 events, CI posts results, the agent does first-pass review, and the merge decision lands in the same room as the evidence. This collapses issue tracker + CI dashboard + chat + code review into one stream. [2]
The caveat, confirmed in Block's own ARCHITECTURE.md: each relay deployment is centralized within its own deployment. Buzz is not peer-to-peer and there is no gossip/replication yet. "Decentralized" describes deployment flexibility (you can self-host), not a peer-to-peer topology. [3]
What this means for you
- You already use Claude Code or Codex on a team: Spend an hour on the desktop app. Wire your existing harness in, build two specialty agents, and run your next sprint of small tasks through Buzz channels instead of pasting context back and forth in Slack. The "see what the agent did, in one signed stream" benefit is the immediate payoff. If you're running Opus 5 specifically, see our Claude Opus 5 effort-dial and pricing guide for the model settings that pair well with a shared Buzz agent.
- You run a small business that's been priced off Slack: Buzz is genuinely free with no user cap, and Apache 2.0 means you can fork it. If your team writes or researches with AI tools, Buzz + one shared LLM subscription is materially cheaper than Slack Pro × N seats + the Slack AI add-on. [4][5]
- You need data sovereignty or you're in a regulated industry: Self-host the relay. The hard path. Only do this once you've validated the workflow on the hosted app.
- You need mobile or you rely on push notifications: Wait. Mobile clients and push notifications are explicitly on the "being wired up" list, not the "works today" list. [2]
FAQ
Q: Is Buzz free to use? A: Yes. The software is free everywhere — hosted by Block, self-hosted from source, or the desktop app. Apache-2.0 license. The only paid piece is the LLM subscription that powers the agent you connect to it (Claude Code, Codex, or Goose via your own API key). [1][2]
Q: Do I need an API key for Claude Code or Codex inside Buzz?
A: Claude Code runs on your existing Claude subscription. Codex needs an OPENAI_API_KEY — notably, not a ChatGPT subscription, the API key (more on the difference in our guide to running OpenAI Codex for free). Goose is the lowest-friction default; buzz-acp with no BUZZ_ACP_AGENT_COMMAND set will spawn goose acp automatically. [3]
Q: Can my whole team share one AI agent through Buzz? A: Yes. The harness runs locally on whoever set it up; teammates tag the agent in from their own devices using their own Buzz identities. There is no platform seat fee. This is the biggest cost win vs Slack. [3][5]
Q: Can I run Claude Code and Codex agents at the same time in one Buzz workspace? A: Yes. Harness and model are per-agent settings, so you can have one agent on Claude Opus and another on GPT-5.6 in the same community, side by side. [3]
Q: Can an agent in Buzz publish content to my website automatically? A: If the harness has your authenticated local session (for example, an existing Claude Code login that has access to your WordPress), the agent can use that access to draft and publish. In the source walkthrough, a Fizz agent drafted a full SEO blog post and dropped it as a WordPress draft for review. Verify the scope of what the harness can reach before giving it production access. [L]
Q: Is Buzz decentralized and self-sovereign like the announcement claims? A: Partially. Each relay deployment is centralized within itself — there is no peer-to-peer replication yet. "Decentralized" describes deployment flexibility (you can self-host the relay on any host) plus the Nostr identity model where keys belong to agents not platforms. Block's own architecture doc confirms this. [3]
Q: Should I self-host the Buzz relay or use the hosted version? A: Start on the hosted app to learn the workflow. Only self-host if you need data sovereignty, regulated-data control, or to run Buzz on infrastructure you own. Self-hosting adds Docker + Hermit and a build step; the setup is real but not for a non-technical team. [2][L]
Discussion
0 comments