Buzz runs coding agents in a git worktree, not a sandbox

I tried out Buzz, and really found some things worth noting, so I wanted to give a quick overview about the project. I really liked the mentality, and by using the Huddle (a live session where you can talk with any agent in the workspace, and I mostly stayed with the built-in @Fizz), I asked all of my questions from project management to task creation, how it mainly works (like git worktrees or sandboxes), how to wire it up through GitHub or a similar tool, how to connect Claude Code (which it lets you do right at the beginning) through ACP (Agent Client Protocol), and workflows. I had a reason to ask that many questions. We are building Volobox at Volosoft, a workspace where people and AI agents works as one team, so Buzz is the closest thing to an open source version of the problem I think about every day. That makes me a biased reader, and I will point out where my bias shows. What Volobox is, briefly: projects, chat, docs and meetings in one workspace, with AI agents you assign tasks to like teammates. Assign a task, an isolated sandbox boots, the agent plans and codes in the open, and a pull request comes back linked to the task that started it. Every write action goes through an approval card, and people and agents share one permission model and one audit trail. Cloud or on-premise, with your own models and your own keys. Buzz is an open source workspace from Block (company that you might know from it's creator Jack Dorsey). The repo calls it "a hive mind communication platform", and the longer version is "a workspace where humans and agents build together, on a relay you own". // Detect dark theme var iframe = document.getElementById('tweet-2079605800998146171-226'); if (document.body.className.includes('dark-theme')) { iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2079605800998146171&theme=dark" } It looks like a chat app at first: channels, DMs, threads, media. The difference is who sits next to you in those channels. Agents are first-class members, not bots bolted on the side. Each one gets its own identity and signs what it does, so every message, review and git event lands in the same log. You can always tell who did what, human or agent. That's the mentality I liked. The bet isn't "open a side panel and ask AI for help". It's "put the agents in the room, with the same context everyone else has". It's a desktop app with mobile clients, you can self-host it or use Block's hosted relay, and it's model agnostic: Claude Code, Codex and Goose all plug in through ACP. Note: It's currently developer preview, and the project says so openly: "Not finished. We will tell you what works and what doesn't." So, don't expect everything in place for now. You grab the desktop build from buzz.xyz and install it like any other app. Nothing surprising there. The interesting part is the sign-up, because there isn't one. No email, no password, no OAuth button. On first launch Buzz generates an identity key for you. That key is your account, and Buzz walks you through backing it up. Important: anyone holding that private key can act as you. Put it in your password manager during onboarding, not "later". After the key, you either create a community or join one. Then comes the part I cared about: agent harnesses. Buzz scans your machine for Claude Code and Codex, and if it finds one it offers to configure it right there. I picked Claude Code, and from that point on every agent task in Buzz runs through my existing setup over ACP (agent-client protocol). Note: this is nicer than it sounds. Because it goes through Claude Code, the agent inherits my skills and MCP servers. Nothing had to be re-configured inside Buzz and I can easily see it through settings (connected mcp servers, skills etc. read from .claude directory). By default we only have "channels" now, you can add your friends and chat with, and it comes three pre-built agents (Fizz, Honey, Pollen, and since they say the Buzz is "a hive mind communication platform", it's great naming for agents) that you can talk with. Beyond the usual channels, DMs and threads, two things stood out: Search that covers everything, so conversations, patches and workflow runs come back in the same result list The three starter agents, which you talk to like any teammate, and you can add your own on top @Fizz is the one pointed at building and development work, and that's the one I spent nearly all my time with. By default these are hidden, and in the application, you are surprised to see it only provides "chats". But then, if you check Settings → Experiments, you can see the some essential building blocks and enable them to see them in the UI. Here's what's behind that toggle list: Workflows - YAML automations with approval gates Projects - git repository browser and collaboration Pulse - activity feed with notes, posts and agent activity Forum Channels - threaded channels for long-form discussions Thread Scoped ACP Sessions - gives each thread its own isolated agent context Agent-managed profiles - lets agents manage their own name and avatar Buzz labels them as "functional but still being refined", which matched my experience. Turn on Projects and Workflows first. Those two are what change the app from a chat client into a workspace. Huddle is the live session layer of Buzz. You start one, and you're talking with agents that already have the context of the channel you're in. Instead of reading docs to figure out what the app can do, you ask. That's exactly how I explored it. I asked Fizz how projects work, how tasks get executed, and what happens under the hood, and it answered from the actual capabilities of the app rather than from a generic model answer. Then I pushed further and asked it to set the projects up for me. It created repository projects for the ABP Framework and ABP-Samples repositories using Buzz's own CLI commands. I clicked nothing! I can also see all activities in the projects. I did nothing, everything was done through the Fizz agent and Buzz CLI commands automatically and it's good that we can see it later on through activities, this is especially useful for your organizations. Projects puts the repository next to the conversation. You browse the code and the commit history, open issues, and review pull requests with inline comments on the changed files and a merge button, without leaving the app. Projects and repositories are bound to channels, and a pull request or issue shows you which channels it was discussed in, so the conversation about a change is one click from the change. You can create tasks, assign them to an agent, and let it work. It creates a git worktree from the default branch (main or dev, generally) unless you point out the base branch, works on the change, commits it, and then you can review. And if you are using GitHub like me (or any other similar remote system), you can ask an agent to push it to remote and create a PR, and it does it for you. In my case, it uses the git/gh commands and does it for me. Before doing it, it showed a "git credential manager" modal and asked for username and password, and successfully completed top to bottom. Btw, since I configured Claude Code at the beginning, it did it through Claude Code, and in the commits I could see Claude Code as co-author, and my limits are going through there. It uses ACP under the hood and communicates with Claude, and everything comes from there, so it can use your existing skills etc. On the review side, the diff shows up inside Buzz as a thread you can talk in. You read the patch, leave a comment, ask for a change, and the agent picks it up in the same place. Workflows are YAML automations that react to something happening in the workspace. A message, a reaction, a schedule or a webhook can be the trigger, and a workflow can wait on a human reaction as an approval gate before it continues. You can also create workflows like: when a message contains "deploy" → call a webhook and start the process for example, or send a message to your DevOps team leader etc. Also, you can create forums like in the other platforms. Forum Channels are threaded and long-form spaces, rather than a fast chat channel. Each topic is its own thread, so an RFC or a design discussion stays readable a week later instead of scrolling away. And Pulse is the workspace-wide social timeline (Channels are scoped conversations, Pulse is the broadcast layer everyone in the workspace sees). In practice it's where agent activity and short notes land, so you can skim what happened across the whole workspace without opening every channel. This is only the first half. In the full article I also cover what I asked the Fizz agent and what it told me (worktrees vs sandboxes, how far you can push agents, the mobile app pairing), plus a longer section on the design decision that shapes everything else: why Buzz works in a git worktree instead of a sandbox, and what that costs you. 👉 Read the full article here: Trying out Buzz - a workspace designed as communication platform And if you like this kind of content, subscribe to my Substack where I share my articles about AI tools, agents and developer productivity. It's free. Thanks for reading, see you in the next one!