Skip to content
Berd

Buzz or Berd

Block makes two agent products, and they are not rivals. Its own description of the pair: start alone, then go multiplayer.

Berd is a private workspace on your own computer, where one person works with agents. Buzz is the shared workspace, where people and agents work together in the same rooms. Most people who use both will use both in the same week, so the question is never "which tool" but "which one for this piece of work, right now".

For how Buzz relates to agent tools in general, see Where Buzz sits in the tool landscape. This page is about the narrower choice between the two Block products.

Where the rest of Berd lives

This guide covers the choice between the two products. Everything about using Berd itself — set-up, shaping a workspace, the built-in agents, automations, safeguards, supporting knowledge and an FAQ — lives in its own companion site, the Berd Coach. Start with Set up Berd if you have not opened it yet, or go straight to the ten built-in agents — Berdy, Agt. Builder, Tinker, Wildcard, Choosey, Pushback, Copycat, Generalist, Builder and Debugger.

The decision

If nobody else needs to see it, add to it, or point back at it later, it belongs in Berd. As soon as one of those becomes true, it belongs in Buzz.

Three separate tests, and the last one catches people out. Work can be entirely your own and still belong in Buzz, because in six months someone may ask how a decision was reached — and the answer needs to exist somewhere other than your laptop.

Berd in brief

An app you install on your own computer for working with AI agents. It gives you one place to work across different models and different harnesses, instead of a separate interface for each. Berd is the environment around the agent, not the agent itself.

The words Berd uses, and what they correspond to in Buzz:

In Berd What it is Closest thing in Buzz
Session, or chat One conversation with an agent A thread in a channel
Project A set of chats grouped around the folders you are working in A channel, roughly
Agent A saved set-up: name, picture, model, instructions An agent — but without its own identity or channel membership
Skill Written instructions the agent picks up when the work calls for them A skill made available to an agent
Automation A saved task you can re-run, on a schedule, with a record of past runs A workflow
Artefact A file from a conversation, opened in a viewer beside it A canvas, or shared media

Ten agents come with it — seven crafted personas with a stated job each (help you use the app, build other agents, push back on a draft, narrow down options, write in your voice, and so on) plus three plain starter agents meant to be grown. They are worth trying before you build your own, because they show what the product is for: an agent shaped around something you do repeatedly, kept, and come back to. The full cast, agent by agent, is on the Berd Coach.

Two limits, by design

Berd stops where the agent's own machinery starts. It can help you with the app around the agent; it does not explain or fix what a particular agent does internally.

Berd has no idea other people exist. No shared rooms, no membership, no identities, nothing. This is not a gap waiting to be filled — it is what the product is. You share by exporting something, not by someone joining you.

Where each one fits

The work Where Why
Reading source material and forming a first view Berd Nobody should see a first pass. There is nothing to keep a record of yet.
Drafting, and reworking the wording Berd It becomes shared when it is ready, not while it is taking shape.
Building an agent or skill around your own way of working Berd This is your own tooling. It goes with you from job to job.
Anything a client or colleague sees or comments on Buzz The moment a second person is involved, the shared record is the point.
Putting a draft to specialists for comment Buzz There is no Berd equivalent of a room where several people respond over days.
Anything you may need to account for later Buzz Every message is signed and kept. A Berd conversation is a file on one laptop.
Work you may hand to a colleague partway through Buzz Context on your computer cannot be picked up by anyone else.

The common mistake is doing solo work in Buzz because it feels tidier — every message then pays for collaboration that is not happening. The opposite mistake costs more: doing the real work in Berd because it is quicker, and finding at handover that the reasoning behind weeks of decisions exists only in your own chat history.

Client material

This is where the choice stops being about convenience.

In Buzz, whoever runs the relay can read everything. That is the established fact the whole hosting decision rests on, and why client or health-sector material needs a relay you run yourself. See Decide your data residency and hosting model.

In Berd, the material sits on your laptop. That removes the relay operator from the picture and replaces them with different questions. Where does the AI provider sit, and what have you agreed with them? What is on that disk if the laptop is lost or stolen? Berd's usage reporting is switched off by default in official versions and excludes your messages, files and credentials — but a modified version may behave differently, so know which one you installed.

Neither is automatically the safer choice. Berd takes one party out of the chain. It also takes out the record, the shared discipline of classifying material, and any chance of a colleague noticing what you put where. A laptop full of unsorted client material is a real exposure — just a quieter one than a hosted relay.

Decide the material first, then the tool

Work out what kind of material a piece of work involves, and let that decide which tool it may go in. Choosing the tool first and then reasoning backwards to whether the material was sensitive is how a rule stops being a rule.

Moving work between them

The two are separate products with separate storage. Until Berd v0.6.3 there was no route between them at all; now there is one, in one direction, and it is narrow by design.

Buzz → Berd: the Buzz Handoff skill. Berd v0.6.3 publishes a skill, buzz-handoff, that lets a Berd agent read a Buzz channel or thread and continue the work privately. You paste a buzz://message link or a channel id into the Berd chat; the agent reads the thread through the Buzz command-line client, summarises only the relevant context, and carries on with you alone. It can also send something back — but only as a reply whose exact text, channel and reply target you have seen and approved, and marked as coming from an agent unless you have set up a separate agent identity for it. Editing the draft is not approval; a change to the text, the channel or the target means a fresh preview and a fresh yes. That is the right shape: the private conversation stays private, and the only thing that reaches the shared record is a message a person signed off.

Two things to know before you rely on it. It needs the Buzz command-line client on your machine with a relay address and a private key in the agent's environment — set up outside the chat, never pasted into it — and the skill refuses to borrow the identity Buzz Desktop manages, so the key you configure is the identity the reply carries. And it is a skill, not an integration: Berd agents that have not been given it do not know Buzz exists. Last checked: Berd v0.6.3 (skills/buzz-handoff/SKILL.md), Buzz v0.5.23 CLI (messages thread --link), 2026-09-10 — read from source, not yet tested end to end. The step-by-step version of this handoff, from the Berd side, is Hand work between Berd and Buzz on the Coach.

Everything else is still manual. Berd can put things on disk: an agent as a shareable card, file or ZIP, a skill as a file, a conversation as an export, and the files themselves. Buzz can take in files posted to a channel, and agents you set up there yourself. Paste the conclusion, upload the file, set the agent up again. The Berd conversation itself does not travel into Buzz; only what you choose to post does.

What follows for now: decide where a piece of work lives before you start it, because moving it later means building it again. If you are unsure whether something will end up involving other people, start it in Buzz. A quiet channel costs little; redoing the work costs a lot.

Explaining the two tools

People are going to ask why there are two.

One is the private workspace where the thinking gets done — much as anyone has always had their own desk and their own way of working that the client does not sit inside. The other is the shared room where the work becomes something people can look at, respond to and rely on later. The line between the tools is the same line that already existed in the work; it has simply become visible.

If a client asks whether their material is on your laptop, that is a fair question with a specific answer, and it is better decided beforehand than in the meeting.

Maturity

Both are early, and neither should be treated as settled.

Berd's source was made public on 2026-08-18 at v0.6.2; v0.6.3 followed on 2026-09-08, a large release that added voice conversation engines, search across every session, agent import from ZIP and the Buzz Handoff skill described above; v0.6.4 (2026-09-10) polished voice and artefact handling and added Pi at onboarding. It is built by a small team at Block, working in the open: anyone can read the source, and bug reports are welcomed, but code contributions from outside the team are not accepted. The Coach's What changed page tracks each release as it lands.

Buzz's state is tracked in the maturity map. The short version is that the shared, collaborative part is dependable and the automation part is not.

One difference worth knowing if automation is your reason for asking: Buzz has a good editor for workflows and an engine you cannot yet rely on for scheduled work or approval steps. Berd has automations with schedules and a record of past runs. Whether Berd's are more reliable in practice is not something this guide has tested — do not assume so, and apply the same rule either way: check that the work actually happened, not that the run finished.

Learn more

Last checked: Buzz v0.5.23, Berd v0.6.4 (release notes and source tree), 2026-09-11.

Reference

Verified against Buzz v0.5.23 · Berd v0.6.4 · Updated 19 Sep 2026