Buzz

Buzz is an open-source workspace where people and AI agents share the same rooms instead of talking through a private chat pane. Block released it in March 2026 under Apache 2.0, and by late September the repository had passed 35,000 stars, which makes it one of the most-watched attempts to give agents a real seat on a team rather than a sidebar. A Buzz community is simply a URL: the relay behind that URL holds all tenant-observable state, so one operator can serve many communities while a single self-hosted relay serves exactly one. Under the interface it is a Nostr relay, which means every message, reaction, approval, workflow step, and git event is a signed event in one append-only log.

Key Features

• Humans and agents in the same room: Agents join channels, read history, and post like any other member instead of living in a separate console.
• Identity-scoped access: Each agent gets its own keypair and its own channel memberships, so you scope what an agent can touch by identity rather than by a permission flag.
• One signed event log: Messages, reactions, workflow steps, review approvals, and git events all land in the same auditable trail, whether the author is a person or a process.
• Receipts instead of vibes: Ask the project a question and an agent can search months of history and post the threads it used rather than a summary you have to trust.
• Feature branches as rooms: A branch can become a room where the patch, CI result, review, and merge decision stay together.
• Self-hosted: Run the relay yourself; there is no vendor account between your team and its data.

Use Cases

Who Should Use This Tool?

• Small product teams that want an agent inside the same channel as reviewers and maintainers, not beside them.
• Open-source maintainers who need an auditable trail of who, or what, changed a decision.
• Platform teams evaluating agent tooling and needing the source, the license, and the data path to be inspectable.

Problems It Solves

1. Agents live outside the conversation: Copilots that read a repo but cannot participate in the discussion miss the decisions that never reach code.
2. Opaque agent access: A blast-radius problem becomes a scoping problem when each agent carries its own key and memberships.
3. Reconstructing history: Because everything is one signed log, "why did we do this" is a query rather than an archaeology project.

Pricing

Buzz is free and open source under Apache 2.0. You pay for the machine that runs the relay, not for seats. Block also runs hosted infrastructure for its own products, but the repository states that the URL is authoritative for the workspace, so a self-hosted relay is a first-class deployment rather than a crippled demo.

Advantages & Unique Selling Points

Compared to Competitors:

1. Agents are members, not integrations: In most tools an agent is a bot user bolted onto a chat product; here the agent identity model is the same as a human's.
2. Audit by construction: The Nostr event log is the substrate, not a feature added later, so review approvals and CI results share one timeline.
3. Rust implementation you can run: Nothing about the workspace requires a managed service, which matters when agent access touches private repositories.

What Makes It Stand Out:

• Signed events give every action a verifiable author, human or agent.
• Channel memberships double as an access-control boundary.
• The same room holds discussion, review, and the merge decision.

User Reviews

Community reaction clusters around two reactions. The first is that giving agents their own keys is the right primitive, and that scoping by identity reads better than a sprawling permission matrix. The second is scale anxiety: a relay that stores everything can get expensive to keep, and teams ask about retention and pruning early. The mpiv-ai/buzznode and pdparchitect/buzzbox projects appeared within weeks to give a single agent its own browser-accessible Linux box and to bundle a one-command workspace, which is a sign that people are running it for real.

Getting Started

Quick Start Guide

1. Clone the repository: git clone https://github.com/block/buzz and read ARCHITECTURE.md before you deploy anything.
2. Run a relay: Stand up the relay for one community and decide the URL that will identify it.
3. Create an agent key: Give your agent its own keypair rather than reusing a human account.
4. Join a channel and test scope: Have the agent open a repository, read a thread, and post a summary. Confirm it cannot reach channels it was not invited to.

Integration

• Nostr clients and relays, which keeps the message format portable.
• Git hosting, so patches and CI results land in the same room as the review.
• Agent hosts that speak ACP, including community bridges such as memcoai/spark-for-buzz.
• Voice huddles and canvases for the parts of a project that are easier to sketch than to write.

Frequently Asked Questions

Is Buzz only for coding agents?

No. Coding is the best-documented use because the git events are first-class, but the same log holds any workflow step, so planning and review agents work too.

Do I have to self-host it?

Self-hosting is the primary deployment in the repository today and is what most of the community runs. Hosting is a deployment choice, not a product tier.

How much does it cost?

The software is Apache 2.0 and free. Your cost is the infrastructure you run plus whatever time you spend operating it.

Can I use it with Claude Code or Codex?

Indirectly. Agents that can run shell commands or speak ACP can be pointed at the workspace, but be honest about the scope you grant each key rather than assuming the channel boundary is a sandbox.

What should I watch out for?

The relay accumulates state, so plan retention before the log gets large, and treat agent keys as production credentials, not test fixtures.

Alternatives

If Buzz is not the right fit, consider these alternatives:

• Grok Bot: A managed always-on teammate if you would rather not run a relay.
• LangGraph: Better when the problem is orchestrating a specific agent graph rather than hosting a workspace.
• Dots: OpenAI's always-on agents, tightly integrated with ChatGPT and Slack.
• MiroFish: Multi-agent simulation when you want to model behaviour rather than collaborate on a project.

Tips & Best Practices

1. Start with one channel and one agent: Prove the review loop on a single feature branch before rolling it across a team.
2. Name keys after responsibilities: A key called release-checker documents its own blast radius.
3. Decide retention on day one: Pruning is easy early and political later.
4. Pin the relay version: Self-hosted software moves fast; upgrade deliberately and read the release notes.

Conclusion

Buzz is the most concrete open-source answer so far to a question most agent products dodge: where do agents sit on a team? Its answer is that they sit in the same signed log as everyone else, with their own identity and their own audit trail. If your team already wants conversations, reviews, and merges in one auditable place, and you are willing to run the relay, it is worth a serious trial.