The modern AI developer's workflow is a chaotic mosaic of browser tabs and terminal windows. One moment a user is prompting Claude for architectural advice, the next they are switching to a local Llama instance for privacy, all while trying to keep project files synced across three different interfaces. This friction—the cognitive load of the context switch—has become the primary bottleneck in agentic workflows. The industry has seen a surge in AI wrappers, but few tools actually address the underlying need for a unified environment where the model, the tool, and the data exist in a single, transparent state.

The Open-Source Blueprint for Agent Orchestration

Block has stepped in to solve this fragmentation by open-sourcing Berd. Released under the permissive Apache 2.0 license, Berd is a desktop application designed to unify AI models, tools, and projects into a single, cohesive workspace. By choosing Apache 2.0, Block ensures that the tool is not just free to use, but free to modify and redistribute for commercial purposes, allowing enterprises to bake their own proprietary workflows into the environment. As of August 18, the project has reached version 0.6.2, bolstered by the contributions of 91 developers on GitHub.

The application is fully cross-platform, supporting macOS, Windows, and Linux, ensuring that the environment remains consistent regardless of the underlying operating system. Within Berd, users can manage chat sessions, attach files, and swap between different agents and models based on the specific requirements of a task. This architecture effectively collapses a fragmented toolchain into a single interface, where project management and automation construction happen in one place. Rather than jumping between separate SaaS platforms, a user can define their project context once and apply various AI agents to it without losing the thread of the conversation or the state of the files involved.

Beyond the Wrapper: Operational State and the Goose Framework

Most AI interfaces are essentially wrappers—thin UI layers that send a prompt to an API and display the response. Berd rejects this minimalism in favor of what is termed operational state. In Berd, the internal machinery is not hidden behind a settings menu or a hidden configuration file; instead, the current project, active files, selected agent, and session state are positioned front-and-center. When a process fails or a model is loading, the system exposes that state explicitly. This transparency allows the user to maintain total visibility into the AI's reasoning and execution path, transforming the AI from a black-box oracle into a manageable piece of software.

This transparency is powered by Goose, a model-agnostic agent framework also developed by Block. Goose leverages the Model Context Protocol (MCP), an open standard that allows AI systems to connect directly to external tools, local files, and third-party data. This means the agent is not limited to generating text; it is an execution entity capable of manipulating data and performing actual commands. Because Goose is model-agnostic, it functions equally well with proprietary frontier models and open-source models running locally, providing a critical hedge against vendor lock-in.

The technical architecture further separates the concerns to ensure performance. The UI is built with Tauri 2 and React 19, a combination chosen to maximize runtime efficiency and keep the desktop footprint lightweight. While the UI handles the high-level management of sessions and contexts, the actual agent loop is handled by Goose, which runs as a bundled sidecar process. These two layers communicate via the Agent Client Protocol (ACP), a standardized communication layer that ensures the UI remains responsive even when the agent is performing heavy computation or complex file operations.

To make these complex roles intuitive, Block introduced Gloopies—animated characters that represent specific agent personas. These are not merely cosmetic; each Gloopie is tied to a specific set of instructions, skills, and tools. For example, the Pushback persona is specifically tuned to challenge drafts and identify logical flaws, while the Choosey persona is designed to narrow down decision parameters. By assigning these personas to visual characters, Berd transforms abstract system prompts into tangible identities, making it easier for users to orchestrate a team of specialized agents.

For those looking to build an independent workspace that does not rely on a specific AI vendor's ecosystem, the Berd approach provides a roadmap. By utilizing an ACP-based separation of UI and runtime and employing lockfiles to fix backend versions, users can maintain direct ownership of their local file access and agent control.

By decoupling the UI from the runtime and embracing open protocols, Block is shifting the power dynamic from the model provider back to the local operator.