SUPPORTING KNOWLEDGE · 0001
What Berd is, and what it refuses to be¶
Berd is Block's open-source desktop app for working with AI agents. It runs on your own computer, holds your chats, projects, agents and skills in one place (and automations, in distributions that enable them — not the v0.6.4 public build), and lets you work across several models and several agent runtimes without switching apps. The agent itself is not Berd; Berd is the room the agent works in.
Brad, who leads AI at Block, describes it as an end-to-end interface for people working through AI tools, with four jobs: make you efficient with those tools, give you access to many models so you can use the right one for the moment, increasingly make that choice for you, and give your work a durable home. The harness underneath, Goose, has not gone away; in his words it is fading into the background.
The thread Berdy, the built-in guide, keeps returning to is that in Berd what you build sticks around: a chat can leave behind a skill or an agent (or, where a distribution enables automations, an automation) that makes the next chat better.
Why it matters to you¶
You are choosing a tool to hold client thinking, your own method and a growing cast of agents. Knowing what Berd is not saves you from expecting the wrong things of it. Four refusals are written into the product:
It will not explain the harness. Berd's own help stops at the app. How a session started, where a setting lives, how to export a chat: in scope. What Goose or Claude or Codex did inside the session, and why: out of scope, by rule. See App vs harness.
It has no idea other people exist. No rooms, no membership, no identities. Sharing means exporting a file. That is the design, not a gap; the shared workspace is Buzz. The Buzz or Berd page carries the one rule for which product a piece of work belongs in.
It is not a chatbot wrapper. Block's product brief lists the anti-references: a generic chat skin, a dashboard of metrics, a marketing site wearing product chrome, friendliness that hides operational state. Loading, compacting, an unavailable provider and a failed import are meant to be shown plainly.
It does not take your code. Berd is built by a small team at Block, in the open. You can read every line, build it and fork it. Pull requests from outside are closed automatically; a reproducible bug report is what they ask for instead.
How to apply it¶
Install it, connect one provider, and give the first chat something real to finish. Berdy's own instructions say the moment Berd clicks is watching a chat do something rather than describe it.
Expect to work in more than one harness. The harness is chosen per session, and two Berd features (voice conversation and steering) depend on Goose. Decide early whether you care about those.
Treat usage reporting as a decision you make once. It is off by default in official builds and excludes your prompts, messages, files and credentials; a modified build may differ.
When something goes wrong, an app problem has three documented reporting routes, of which the public build has one, the GitHub issue; a harness problem goes to the harness.
If you ignore this¶
You will file a Goose behaviour as a Berd bug and get nowhere. You will do client-facing work in Berd because it is quicker and discover at handover that the reasoning lives only in your own chat history. Or you will wait for a multiplayer feature that is not coming, because that is the other product.
Examples¶
Your own desk, in your own office, with your own way of working. The client never sits inside it. When something is ready for other people, you take it to the shared room.
A consultant reads three source documents with Generalist, drafts a first position with Copycat in their own voice, and only then posts the conclusion to the engagement channel in Buzz. Nothing of the first pass leaves the laptop.
Asking Berd's help why a Claude session refused a tool call. Berd will tell you that is the harness's business, and it will be right.
Do it: Install Berd and get through onboarding
Last checked: Berd v0.6.4, 2026-09-11.
Reference¶
- block/berd — README.md
- block/berd — PRODUCT.md
- block/berd — distro/skills/berd-help/SKILL.md — the scope rule
- block/berd — TELEMETRY.md
- block/berd — CONTRIBUTING.md
- Brad, Block's AI lead, on Block's On The Block (September 2026) — Berd's four jobs, and the harness fading into the background
- Buzz or Berd — the Buzz Guide's page on the two products
- Verified
- Berd v0.6.4 · 2026-09-11