Block Just Open-Sourced Their Internal AI Command Center — And It Unifies Every Coding Agent

Block — the fintech giant behind Square and Cash App — just open-sourced the desktop app their own teams use daily to wrangle AI coding agents. It’s called Berd. It runs on your machine. It speaks Agent Client Protocol. And it might be the first tool that makes working with multiple agents feel like a coherent workflow instead of a fragmented mess.

Released August 18, 2026. Apache 2.0 license. 714 stars and climbing.

The Problem Was Never the Models

By mid-2026, the model layer was solved. You had Claude Code for reasoning depth. Codex for parallel throughput. Goose for extensibility. Each one capable. Each one living in its own terminal window, with its own config, its own context, its own way of doing things.

Block’s engineers lived this fragmentation. They’d switch between Claude Code for architecture, Codex for bulk refactors, Goose for custom tooling — copying context manually, losing track of which agent knew what, rebuilding project state every session.

“We didn’t need another model or agent harness. We needed a consistent environment around them.”

That’s the opening thesis from Block’s own design writeup. Berd is that environment.

What Berd Actually Does

Berd is a native desktop application (Tauri, TypeScript/Rust) that wraps your agents in a unified interface. Think of it as a command post, not another agent.

Supported agents at launch:

• Claude Code (Opus, Sonnet, Haiku)
• OpenAI Codex
• Goose (Block’s own framework, now under the Agentic AI Foundation)
• GitHub Copilot
• Amp Code

The unification layer is ACP — Agent Client Protocol. Berd doesn’t fork or wrap the agent binaries. It connects to them via the same protocol Anthropic, OpenAI, and Block standardized. When Claude Code adds a feature, Berd gets it automatically. No plugin lag.

Persistent Projects, Not Throwaway Chats

The core abstraction is the Project. A Project in Berd holds:

• Conversation history — full, searchable, persistent across sessions
• File context — folders you’ve pinned, files you’ve referenced
• Agent roster — which agents are available, with what roles and skills
• Instructions — AGENTS.md-style guidance that travels with the project
• Skills & tools — reusable capabilities you’ve built or installed

You don’t “start a new chat.” You open a project. The context is already there. The agents already know their roles. You pick up where you left off.

Agents With Character (Literally)

Block’s design team asked: How do you make an abstract AI agent easier to understand?

Their answer: Give it a face.

Berd ships with animated avatar collections — the flagship “Gloopies” — that map to agent configurations. A code-review agent looks different from a test-writing agent. A travel-planning agent has its own visual identity.

This isn’t cosmetic whimsy. In a workspace with five active agents, instant visual recognition beats reading role labels every time. The avatar is the configuration summary.

“The avatars make the agent recognizable. Its role, skills, and tools make it useful.”

You create custom agents by chatting with Berd. Describe what you want. It generates the role, skills, tools, and avatar. No YAML editing required.

Built on Goose, Connected to Buzz

Berd didn’t appear from nowhere. It grew from Goose — Block’s open-source agent framework (January 2025, now 17k+ stars). In December 2025, Block contributed Goose to the Agentic AI Foundation under the Linux Foundation, alongside MCP and AGENTS.md.

The architecture is clean:

• Goose = agent runtime, tool execution, model orchestration
• Berd = desktop shell, projects, context, UI, configuration
• ACP = the protocol between them

This separation means Berd can support any ACP-compliant agent. Goose handles the loop. Berd handles the human.

And when solo work becomes team work? That’s Buzz — Block’s newly released collaborative workspace where people and agents share rooms, identities, and a searchable relay. Berd is the private space. Buzz is the multiplayer layer. Same primitives, different scope.

The Numbers That Matter

Critical detail: Block is not accepting external PRs right now. They’re sharing the codebase, not outsourcing maintenance. The repo is a reference implementation — “here’s how we do it” — not a community governance experiment.

Why This Matters for You

You probably won’t run Berd in production at a fintech scale. But the pattern is worth stealing:

1. One home base for every agent — stop alt-tabbing between terminals
2. Persistent context per project — stop re-explaining your codebase
3. Visible agent configuration — stop guessing which tools are active
4. Reusable skills across agents — stop rewriting the same prompts
5. ACP as the universal connector — stop waiting for vendor-specific plugins

Block proved this works inside a company that moves real money with real compliance requirements. The tool is battle-tested. The protocol is open. The code is yours.

How to Try It

# Download from GitHub Releases (macOS, Windows, Linux)
# Or build from source:
git clone https://github.com/block/berd
cd berd
# Requires: Rust, Node.js, pnpm
just bundle  # stages Goose backend, builds Tauri app

Requirements: Goose backend (bundled), your API keys for Claude/Codex/Copilot, a project folder.

The Honest Take

Berd v0.6.2 is early. Mac/Windows/Linux builds exist but the experience will have rough edges. The agent roster is hardcoded for now. Custom agent creation works but the skill ecosystem is nascent. Block explicitly says: this is how we work, not a finished product for everyone.

But the direction is unambiguous. The industry has spent two years optimizing models. The next two years will be spent on orchestration, context, and human-agent interface. Berd is one of the first serious attempts at that layer — from a company that actually uses it.

“Start alone, then go multiplayer.”

That’s the arc. Berd is the alone part. It’s free, open, and available today.

Related Reading

• Block’s design writeup on Berd
• Goose framework (Agentic AI Foundation)
• Agent Client Protocol specification
• Buzz — Block’s collaborative agent workspace
• Our comparison: Goose vs Claude Code extensibility
• Complete guide to AI coding agents 2026

Related articles

• Claude Code Alternatives in 2026: 12 Options Compared
• Cline Desktop Grew Up Fast — Heres What v0.0.24 Actually Fixes
• Claude Code: Skills vs Subagents vs MCP — The 2026 Decision Guide

Disclosure: This article covers an open-source release from Block (NYSE: SQ). No affiliate relationships exist. terminalblog.com is an independent publication.