If you have just typed "buzz jack dorsey" into a search bar, here is the answer in one paragraph. Buzz is a brand-new, completely free app from Jack Dorsey that turns your AI agents into a team of employees — agents that talk to each other, hand work between themselves and fix each other's mistakes inside one group chat. Julian Goldie installed it, built a full agent team in around 20 minutes, and watched it write, quality-check and publish a blog post with barely any input from him. This is the full story: what Buzz is, how that build went, and where it still falls short.
Why Everyone Is Searching Buzz Jack Dorsey Right Now
Two things collided. The first is the name: Jack Dorsey co-founded Twitter and is the man behind Block, so a new app from him travels fast. The second is the pitch, which is unusually concrete for an AI product: stop juggling separate chat windows and give your agents one shared workspace — a group chat that looks like Slack — where they behave like colleagues rather than tools.
Buzz is free and open source. The only money involved is whatever you already pay for the model subscriptions you plug in, such as Claude Code or Codex. You can download it as a desktop app, which is the easy route, or grab it from GitHub, which is the more complicated one. It is also very young — roughly seven days old when Julian tested it — and that youth shows in ways covered further down.
What Buzz Actually Does
The core idea: your AI agents become employees in a shared chat. One message from you can set off a chain of work — one agent writes a blog post, a second designs images for it, and a third checks everything and corrects the mistakes it finds.
The detail most write-ups miss is context. Because Buzz runs on your existing Claude Code setup, it inherits everything you have already connected. In Julian's build, his agent read his Obsidian vault and suggested personalised automation ideas, pulled keyword data from his Google Search Console account, and published a fully formatted post straight to WordPress without Julian logging in to anything.
You are not starting from a blank assistant. You are giving a team chat to the tools and data your agents already know.
The 20-Minute Setup
Julian's whole system took about 20 minutes:
- Install the app and sign in. The desktop app is the easiest path; GitHub is there if you like to tinker.
- Install the CLIs. Buzz sits on top of command-line agents: Claude Code, Codex and Goose. Goose comes from Block, Jack Dorsey's company, and is downloaded separately, then connected as a provider.
- Pick your default agent. In settings, under agents, choose the default agent and model. Julian kept Claude Code as his default.
Two caveats from his testing. Goose running free models did not work so well for him. And when he asked Fable 5 to help configure Goose, it messed up the setup — his advice is a blunt don't.
The settings menu goes surprisingly deep for week-one software: appearance, notifications, hotkeys and shortcuts, multiple agents (Codex, Grok build, Claude Code, Goose and the Buzz agent) and experimental features including workflows, projects, polls, forms and agent-managed profiles.
📺 Watch: Jack Dorsey's Buzz + Claude Code is INSANE!
Channels and Custom Agents
Work lives in channels, one per workflow. Julian created a second brain channel wired into his Obsidian vault, an image generation channel and an SEO blogs channel.
Custom agents are where it gets interesting. You write a plain-language description of the job, pick a provider, and add an emoji or image avatar. Julian built an Image Designer told to produce multiple variations of everything it makes, and an SEO agent instructed to use his SEO skill, find keywords and publish to WordPress.
Agent Teams: The Part That Feels Like the Future
Teams are the headline feature. Buzz lets you group agents under a named team with shared instructions — Julian combined his Image Designer, his SEO agent and a Quality Controller he created into a marketing team. From then on he tags the team rather than the individuals, and the agents work autonomously, tagging each other as needed.
Here is the real exchange from his build. The SEO agent drafted a post. The Quality Controller reviewed it and caught that the reported word count was wrong — 1,970 claimed against 1,580 actual — along with capitalisation errors. It told the SEO agent to fix both, checked the corrections, confirmed them, and only then was the post published. That is a manager-and-employee loop running between two AIs, with a paper trail you can read line by line.
If you want an AI agent team making you money, check out the AI Profit Boardroom — Julian's community for building exactly these systems, with his agent builds, prompts and step-by-step walkthroughs inside. → See the agent systems members are running
📺 Watch: Jack Dorsey's Buzz DESTROYS Hermes Agent?
Buzz at a Glance
| Question | Answer |
|---|---|
| Price | Free and open source; you only pay for the model subscriptions you connect (Claude Code, Codex) |
| Who is behind it | Jack Dorsey; the Goose CLI comes from his company, Block |
| How to get it | Desktop app (easiest) or GitHub (more complicated) |
| Agents | Claude Code, Codex, Goose, plus custom agents you describe yourself |
| Teams | Group agents under shared instructions and tag the team once |
| Oversight | Step-by-step activity view, reply timer, six months of searchable history |
| Permissions | Each agent has its own keys, channel memberships and audit trail |
| Humans | Invite people via link; they can bring their own agents with limited access |
| Maturity | Around a week old and experimental; mobile app on the way |
Oversight, Permissions and Inviting Real People
Visibility is handled well. You can open any agent's activity and see every step and tool it used, there is a reply timer, and history stays searchable for six months.
Permissions are siloed sensibly too. Each agent carries its own keys, its own channel memberships and its own audit trail, so you can fence off what each agent or project is allowed to touch.
And Buzz is not agents-only. You can invite humans into the community via a link; they can bring their own agents, and you control what they access. A share-machine option even lets members run agents through remote models — Gemma 4 is one example, though Julian has not tested that route yet.
📺 Watch: Hermes Agent + Buzz is INSANE!
The Honest Limits
In his Goldie Bench notes — Julian's own hands-on testing, not an objective scoreboard — four limits stood out:
- Agents only act when tagged. Nothing happens until you, or another agent, mentions them.
- Big tasks can look stalled. Long jobs sometimes appear to time out; a nudge usually restarts them, and model choice matters — Opus 5 in particular can be slow.
- It is about seven days old. Buzz is experimental software, and it behaves like it in places.
- No mobile app quite yet. One is coming, designed to connect to your desktop, but it was not available at the time of testing.
Where Buzz Fits in Julian's Stack
The obvious question: how does this sit next to Agent OS? They solve different layers. Buzz is the most intuitive way Julian has found to run team-style agent orchestration — multiple agents collaborating in a chat anyone can understand. Agent OS and Hermes go deeper: always-on, autonomous business systems that keep working whether or not a chat window is open.
They are converging rather than competing. Julian has already set Fable 5 to work on connecting Hermes into Buzz, so the deep system and the friendly chat layer can talk to each other.
FAQ
Is Buzz really free?
Yes. The app is free and open source. Your only costs are the model subscriptions you plug in, such as Claude Code or Codex.
Do I need to be technical?
Not especially. The app route is install, sign in, add your CLIs and pick a default agent — Julian went from download to a working three-agent team in about 20 minutes. The GitHub route is the more complicated option.
Can other people join my workspace?
Yes. You invite humans via a link; they can bring their own agents, and you decide what each person and agent is allowed to access.
Which model works best inside Buzz?
In Julian's testing, Claude Code as the default was the reliable choice. Goose with free models underperformed for him, and Opus 5 could be slow on bigger tasks — one tester's experience, not a verdict.
The Bottom Line
Buzz is the clearest sign yet of where agent software is heading: away from single chat windows, towards teams of specialised agents that brief, check and correct each other. It is free, a working marketing team took Julian around 20 minutes to build, and watching one agent catch another's wrong word count and force a fix is genuinely impressive.
It is also a week old, experimental and occasionally needs a nudge. Install it, build one small team, and see for yourself — then, when you want that team running a business rather than a demo, you know where the deeper systems live.
