Block Berd agent workspace is open-sourced under Apache 2.0

Block has open-sourced Berd, a locally installed desktop application for organizing work with AI agents across configured models, providers and harnesses. The Block Berd agent workspace is available under the Apache 2.0 license for macOS, Windows and Linux, giving users permission to use, modify and redistribute the software, including commercially, according to VentureBeat.

Berd is a workspace and orchestration layer, not a foundation model or a new agent runtime. That distinction is the practical point for teams assessing it: the app brings projects, sessions, context and configuration into one desktop interface, while the underlying agent work is handled elsewhere.

What is Block's Berd agent workspace?

Block initially developed Berd for employees working across its Goose framework, Anthropic's Claude Code and OpenAI's Codex, where interfaces, configuration and context management differed, VentureBeat reported. The public application is intended to give users one environment around whichever providers and harnesses they configure, rather than a Block-selected model stack.

The workflow includes chats, file and folder attachments, persistent projects, agent and model selection, provider configuration, skills and extensions, session history, and automations. Persistent projects are meant to retain a working set of files, instructions and agent configurations, so users can return to the same project without rebuilding that setup for each task.

Block says the desktop form factor is deliberate. Brad Axen, the company's head of AI Capabilities, told VentureBeat that the application is designed to work directly with local projects, files, tools, repositories and agents running on or connected to the user's computer.

How does Berd differ from Goose?

Berd is built with Tauri 2 and React 19 and communicates with Goose through the Agent Client Protocol, according to VentureBeat. Goose runs as a bundled sidecar process and handles the agent loop. Berd manages the surrounding work surface: projects, sessions, context, agents and configuration.

That division means local storage should not be confused with fully local AI processing. Welcome.ai reported that Berd uses a local-first model in which conversation histories and credentials are held on the user's device. A configuration can still send work to remote models, depending on the provider a user connects. Teams should therefore assess the data handling and terms of each configured provider alongside Berd's local application design.

Berd's build process also pins a specified Goose backend version, verifies the cached binary against that version and packages it with the application, VentureBeat reported. Developers can substitute a different Goose binary for local testing. The approach provides a defined default backend version instead of automatically taking whichever Goose build is current.

Why does Berd give agents avatars?

Block has also given agents visual identities, including animated characters, rather than presenting each as an undifferentiated chat window. The company says the avatars make agents recognizable; their roles, skills and tools are what make them useful. This is a design choice for distinguishing configured agents, not evidence of a measured productivity or security outcome.

Berd arrives as agent developers increasingly package model access, tools and workflows into more structured work surfaces. Its open-source position is narrower than a claim of universal interoperability: the reported capability is to configure preferred providers and harnesses, with Goose serving as the bundled backend. For operators comparing models for defined tasks, the relevant question remains whether each connected provider performs acceptably in the workflow, a question addressed in AI model evaluation.

This story draws on original reporting from VentureBeat.