Published: 2026-08-02

Buzz: a Slack-style workspace where every member is an AI agent

Chapters / key moments (click to jump — plays here on the page)

Buzz looks like Slack — channels, DMs, threads, emoji reactions — but the members in the sidebar are AI agents rather than people, each with its own identity and a record of what it did. FuturMinds installs it, wires it to existing Claude Code and Codex sign-ins, and builds a three-agent content team (lead, researcher, writer) that turns a one-sentence brief into a finished draft. The verdict at the end is deliberately unhyped: promising, but not a Slack replacement yet.

Source video

"Jack Dorsey's Slack for AI Agents Just Dropped" by FuturMinds — Watch on YouTube →

Key Takeaways

Agents are first-class members, not bots in a channel. Each agent has its own name, its own identity, and an activity log, so you can always trace which agent did what. They can be bundled into teams that hand work off between themselves.
It runs on harnesses you already pay for. Buzz uses Claude Code or Codex as the agent "harness" and picks up desktop accounts that are already signed in. There is no separate model billing — you choose a default harness and model during setup, and can assign a stronger model to the lead agent and smaller ones to sub-agents.
"Decentralized" here means data ownership, not crypto. The video addresses this directly: there is no coin, no token, and nothing to buy. It means you host and own your own data.
MCP servers are inherited, not configured in-app. Buzz has no UI for connecting MCP servers. Anything you want agents to reach — Notion, Airtable — gets added to Claude Code or Codex instead, and the agents inherit it through the harness.
Per-agent invocation permissions. Advanced settings control who may send instructions to a given agent — just you, or anyone on the team. Useful for keeping an expensive or destructive agent from being triggered by everyone.
The three-agent demo worked from one message. Tagging the team with a blog-post brief auto-added the agents to the channel; the lead broke the work into tasks, tagged the researcher and writer, the researcher returned six sourced points, and the writer produced the draft. Progress is visible through status emojis and a "view activity" plan view.
Deployment options are broad. The project is open source with roughly 20K GitHub stars at time of recording. It can run locally via a desktop download from buzz.xyz, or be hosted on a cloud server, deployed to Railway, or installed on Hostinger in one click.

Gotchas & Caveats

Back up your identity key immediately

Setup generates an identity key and there is no password reset. If you lose it, the account is gone. Back it up before doing anything else in the app.

Not a Slack replacement today. The creator's explicit verdict — it has bugs and is an early version. He advises against migrating a business off Slack until it matures.
Approvals are half-built. The approval flow is incomplete in this release.
No agent memory or knowledge documents. There is no way to manage what an agent remembers or to hand it knowledge docs, though agents do accumulate memory within a channel over time.
No session or context controls. You cannot cap the number of sessions or set the context window size.

What you can actually set up from this

Extracted from the video's own transcript — the specifics the original summary left out.

Configuration shown

Lead agent instruction

When I give you a topic, break it into steps. Assign the research to Researcher and the writing to Writer. Then review what comes back and post the final draft.

Template shown in the video. Adapt to your project — do not copy verbatim without reviewing each line.

Reproducible steps

Install Buzz
Buzz is free and open source (about 20k GitHub stars at the time). Run it locally by getting the app from buzz.xyz, or host it on a cloud server: deploy on Railway or use the one-click install on Hostinger, about five minutes. It runs your agents on Claude Code or Codex using your existing accounts.
Create your identity key and back it up
First launch asks you to create a new identity key. There is no password reset, so back it up somewhere safe before doing anything else.
Pick the agent harness
The setup shows Claude Code and Codex as connected if you are signed in to both on the desktop. He chose Claude Code as the default harness with the Sonnet model.
Create a community and channels
Create a community (he named his Content), which opens Builderlab in the browser to connect your Buzz identity with a copy-code step. Then create channels by giving a name, channel type and visibility, and add team members or agents.
Create three agents and a team
In the Agents tab create Lead, Researcher and Writer. Under advanced settings you choose who can send instructions to an agent: only you, or anyone on your team. Create an agent team (he called it Content Team) with a description and team instructions and add the three agents.
Give the team one brief
Tag the team in a channel with a one-line task (his: a blog post on why AI agent teams beat single agents). The agents are added to the channel, show working emoji, and View activity expands the plan. Lead splits the work, Researcher posts six points with sources, and Writer produces the draft.
Add tools through your harness
Buzz has no place to connect MCP servers. To give agents access to Notion or Airtable, add those MCP servers in Claude Code (or whichever harness you use) and the agents can use them. You can pick a stronger model for the lead and a smaller one for the sub-agents.

Gotchas

He does not call it a Slack replacement yet: it has a few bugs, approvals are half built, there is no way to manage agent memories or give agents knowledge documents, and no control over the number of sessions or the context window size.
'Decentralized' here means you own your data; there is no coin or token.
Losing the identity key means losing the account.