Buzz ai - what it is and overcoming installation issues on Windows

I am a member of a Skool group called Claude At Work, which runs daily Google Meet sessions where you get to speak to three different business/Claude experts and get help with any projects you are running. It was in this group that I kept seeing the word Buzz, and eventually I asked about it and was immediately excited to try it. Why?

I wanted to look at a different way of working with AI that goes beyond just using AI-assisted coding to build software, and includes more automation and systems that could do the work for me. I asked in the Claude At Work group whether Buzz could help me with my Content Planning software. Rather than starting my computer and asking for the topics of the day, I wanted to start my computer and find the content already created and waiting for me. The answer was yes.

So I decided to install it.

What is Buzz?

Buzz is an open-source project from a company called Block and is designed to put people and AI agents into a shared workspace, an almost Slack-type environment. In this environment you can ask your agent to get work done, and you can automate workflows so they do work without you, but that need your approval. This is humans and AI agents working together.

You are probably familiar with opening an AI assistant, asking a question and receiving an answer, which works well for individual tasks but can feel disconnected from the whole project. I tend to ask questions about what I am thinking about doing in one chat, then start a new chat to get an outline or code, and when something goes wrong that is another new chat. So it all becomes a collection of chats, and when I go to search for something my experience is that it can be a nightmare, especially if what you are looking for is in the middle of a larger chat and it's 3 months later and you can't remember the original question.

With Buzz all of that information exists together in one group, maybe under a specific channel, and you may even have dedicated the task to one agent. Let's call her Henri. So now, depending on how I set Henri up, I can ask her to find out information on a bug that showed up two months ago and what we did to fix it, or ask her to find that specific chat I mentioned earlier that I was looking for, without having to go through everything myself.

What can you do with Buzz?

Buzz is a developing project, but its current capabilities extend beyond basic messaging.

Sort conversations into channels. Channels give you places to discuss particular projects, topics or tasks. Threads help keep related conversations together.

Bring AI agents into discussions. Compatible agents can be connected to Buzz and respond when they are mentioned. This gives you a way to involve an agent in an ongoing project rather than treating every interaction as a separate session.

Keep project information together. Buzz supports searchable conversations and project-related activity. The aim is to make it easier to find previous discussions, decisions and relevant information.

Work with development activity. The project includes Git-related functionality, allowing development events and code-related discussions to become part of the wider working environment.

Use automated workflows. Buzz supports YAML-based workflows with triggers such as messages, reactions, schedules and webhooks. These can be used to connect events to repeatable actions.

How does Buzz connect to Claude Code and other AI agents?

One of the features that interested me was the ability to connect coding agents.

Buzz supports agents that use the Agent Client Protocol, or ACP. This is a protocol that allows compatible applications and AI agents to communicate through a common interface.

The Buzz project documents support for agents including Claude Code, OpenAI Codex and Goose. The precise setup differs depending on which agent you want to use.

Once connected and configured, an agent can listen for messages directed to it in the channels it belongs to and respond through Buzz.

This is different from merely opening Claude Code in a terminal and working on a project independently. The idea is to make the agent available as part of a shared working environment.

There are, however, some important practical requirements. Installing the Buzz desktop application does not necessarily mean every agent will be ready to use immediately. The relay, agent configuration, permissions and any required adapters all matter.

What you need to install Buzz

There are two ways to get Buzz, and what you need depends on which one you choose. The packaged installer is the quicker route. Building from source is for people who want to run the whole stack, relay included, on their own computer. I used the desktop app, then needed Node.js 24 for the npm command that fixed my problem.

My experience installing Buzz on Windows 11

I installed Buzz Desktop on my Windows 11 computer to see how it worked in practice. I already had Claude Code installed and can run it from PowerShell (I watched YouTube videos), so I thought it would be easy. However, during Buzz setup, Claude was reported as not installed, and I know it is and it's working.

The actual problem

During the initial setup, Buzz displayed a screen asking me to connect an AI subscription through a compatible tool. As you can see from the image, Buzz showed Claude as “Not installed” and could not detect it. So with some help from ChatGPT (I save my Claude Code tokens for the very important stuff) I used PowerShell to investigate the problem.

The key discovery was that Claude Code itself was working correctly, but the separate Claude Agent Client Protocol (ACP) adapter was not installed.

Fixing the issue

claude --version
where.exe claude-agent-acp
npm list -g @agentclientprotocol/claude-agent-acp
npm install -g @agentclientprotocol/claude-agent-acp

Final result

Buzz AI is working

• Claude Code is connected.
• Buzz has completed its initial setup and I have my API key to log in.
• The welcome screen is displayed and I started a new group.

The important lesson

Installing Claude Code alone was not sufficient for this Buzz setup. The missing ACP adapter was the additional component identified and installed. Restarting the computer afterwards allowed Buzz to detect the setup correctly.

If I encounter the same issue again, I should check the ACP adapter installation before reinstalling Claude Code or changing my Windows PATH.

How I set up my agents and what I changed

Once Buzz was working, I joined a community that Buzz hosts and got three starter agents called Honey, Pollen and Fizz. Each one has a short personality written into its instructions. Honey helps with writing, Fizz turns ideas into action and Pollen does research. Each agent is a Claude Code session that Buzz starts through the ACP adapter, so all three use my Claude subscription.

Two details surprised me. The relay for my community is hosted by Buzz, so my channels and messages sit with the hosted relay, not on my computer. The agents run on my own PC. When my PC is switched off, the agents stop. They also only act when I mention them in a message, and they only take instructions from me, so nothing gets published or changed unless I ask.

Why ten workers per agent is not a great way to work

I asked Claude Code to read the agent logs, and each agent was starting 10 workers. A worker is a separate Claude Code process waiting for a message, and 10 is the app default. With three agents that was about 30 background processes on my PC, all sitting idle most of the day.

For one person that is far more than needed. Buzz deals with one request at a time in each channel, so extra workers only help when several channels are busy at once. The rest are programs running on my computer for no reason, and if they all started work together they would use up my plan limits faster.

Changing it to 2

The setting is called Parallelism. Open the agent from the Agents page, choose Edit, expand Advanced and type a number. Leave it blank and Buzz uses its default of 10. I typed 2 for each agent.

There was a catch. Saving in that screen changed the agent template, but the copies already running in my channel kept 10, and the logs still said so after a restart. To fix it I closed Buzz, made a backup of managed-agents.json, changed the parallelism value from 10 to 2 on the three running agents and opened Buzz again.

To check it, look in the agent logs folder at %APPDATA%\xyz.block.buzz.app\agents\logs for the line that starts with buzz-acp starting. It now says agents=2. My PC went from about 30 background processes to 6, which is 2 per agent.

When you might need more than 2

Two works for me because I am the only person using these agents. You may want more if:

• Several people message the same agent at the same time.
• You run many channels at once, for example one per website, and agents are busy in more than one.
• Scheduled workflows start several jobs together, such as drafting for every site at the same hour.

Raise the number one step at a time, to 3 or 4, and watch how quickly you use your plan limits. Every extra worker is another Claude Code session drawing on the same subscription.

Letting Claude use my computer to ask the agents questions

Buzz agents answer only when someone writes to them inside the Buzz app. To test mine, I wanted Claude to ask them questions, so I let it use my computer. The Claude desktop app has a feature called computer use. Claude takes screenshots and uses the mouse and keyboard in the apps you allow.

It asks permission for each app before it touches anything. I allowed Buzz first. Buzz was hidden behind other windows, so I also allowed File Explorer, which let Claude click the taskbar to bring Buzz forward. The permissions have tiers. Browsers are view only, terminals allow clicks but no typing, and File Explorer allowed clicking only. Any app you have not allowed is blanked out in the screenshots it takes.

With that in place, Claude opened the Agents page, changed the settings, and typed a message to Honey asking her to list my websites from the Content Planning Studio. Then it read her reply.

• Allow only the apps the job needs, and add another only when you are asked.
• Stay at your computer and watch the screen while it works. You can stop it at any point.
• Expect small mistakes. A mention did not register a couple of times, and a few stray messages ended up in my channel for me to tidy up.

Once the agents were running properly, I connected them to my Content Planning Studio with read-only access, so they can read my content plans and assets but cannot change them.

Is Buzz free and open source?

Yes, Buzz is an open-source project, and its source code is publicly available on GitHub. You can explore the code, read the documentation and check the available desktop releases. The project also documents ways to run your own relay or use a relay hosted elsewhere.

One detail to understand is that the desktop application and the relay serve different roles. The desktop application is the interface you use, while the relay provides the shared infrastructure that connects participants and stores workspace events.

For that reason, installing Buzz Desktop is not necessarily the whole setup. The default connection expects a relay at localhost:3000, unless you configure another one or use an available hosted instance.

Windows users should also check the release information carefully. The project has distributed Windows installers marked as unsigned alpha releases, so it is sensible to review the current release notes and installation guidance before proceeding. You can find the project here: Buzz on GitHub.

Is Buzz worth exploring?

What interests me about Buzz is the idea of treating AI agents as participants in a working environment rather than as separate tools that you open only when you need an answer.

As people begin using multiple AI agents for coding, research and other tasks, keeping conversations, decisions and results connected becomes increasingly important. A shared workspace offers one possible way to approach that problem.

Buzz is still evolving, so it is important to distinguish its current features from the longer-term vision. It will not automatically make every workflow simpler, and connecting agents requires the appropriate configuration.

For me, installing it and getting Claude Code connected was a useful first step in understanding how the system works. I am interested in seeing how this kind of workspace develops and what practical uses emerge as the software matures. As I begin to connect it directly to my projects and get them working, that is where I will really see the true power of what it can do.

If you are curious about working with AI agents in a more integrated way, Buzz is worth a look. Start with the official Buzz repository, review the documentation and release notes, and decide whether its approach fits the way you work.