Block Unveils Berd: The Open-Source Multi-Model Desktop Workspace for AI Agents 🤖

Block has open-sourced Berd, a local-first desktop application engineered to streamline how professionals interact with autonomous AI agents across different models, frameworks, and projects. Released under the permissive Apache 2.0 licence, Berd provides a unified workspace for macOS, Windows, and Linux. By running natively on the desktop rather than inside a web browser, the application connects directly with local files, code repositories, terminal tools, and on-device agent runtimes.

The Daily Work Surface for Multi-Agent Workflows

As modern organisations adopt specialised AI tools—ranging from OpenAI’s Codex to Anthropic’s Claude Code and Block’s own Goose framework—managing multiple agents often results in fragmented context and scattered configurations. Berd addresses this challenge by serving as an orchestration layer that keeps operational state clear and visible.

Rather than hiding settings in administrative menus, Berd explicitly displays which project, local files, active tools, and provider models are shaping a given conversation.

• Persistent Projects: Users can establish dedicated project environments with specific file collections, custom instructions, and agent configurations, avoiding the need to reconstruct context for every task.
• Model Agnostic Flexibility: The workspace allows teams to pair specific jobs with appropriate models—from top-tier proprietary APIs to lightweight, locally hosted open-weight models using tools like Ollama.
• Accessible Beyond Engineering: While powerful enough for complex software development, the interface is designed to let non-technical users begin with a simple chat and progressively add structured tools and context as needed.

Distinct Visual Identities and Personas

To solve the common issue of navigating generic, identical chat windows, Berd introduces distinct visual identities for different agent configurations. Using animated characters known as "Gloopies", the platform translates invisible background instructions, permissions, and tool access into recognisable personas.

Predefined personas showcase how these visual roles support specialised tasks:

• Pushback: Functions as a critical editor, challenging assumptions and identifying weaknesses in drafts.
• Choosey: Helps users evaluate competing options and structure complex decisions.
• Copycat: Adapts to write and format content in the user's specific tone and style.
• Tinker & Wildcard: Tailored for experimental workflows, debugging, and open-ended tool usage.

These visual identities ensure users can instantly identify an agent's capabilities and assigned guardrails before initiating a workspace session.

Architecture Built for Performance and Control

Berd operates purely as an orchestration layer rather than a custom foundation model or execution engine. The desktop interface is built using Tauri 2 and React 19, delivering a lightweight footprint.

Behind the scenes, Berd communicates via the Agent Client Protocol (ACP) with Goose, the open-source agent framework Block contributed to the Agentic AI Foundation. Goose runs quietly alongside Berd as a bundled sidecar process, managing the underlying execution loop between models, files, and external systems via open standards like the Model Context Protocol (MCP).

To maintain technical stability, Berd’s build process pins specific Goose backend versions using a lockfile and verifies binary integrity. This strict boundary prevents unexpected backend changes while still allowing developers to substitute custom builds for testing.

Local-First Data Privacy and Enterprise Customisation

Data governance and endpoint security are central to Berd's architectural design. The application prioritises user control through several structural safeguards:

• Local Data Storage: Conversation histories and session logs are retained locally on the user's device within the internal session database.
• Secure Credentials: API keys and provider credentials are stored inside the operating system's native keychain.
• Privacy-Focused Telemetry: Usage telemetry is completely disabled by default. When explicitly enabled by the user, the transmitted data excludes prompts, file contents, code, and personal identifiers.
• Enterprise Distribution Seams: Companies can create custom internal distributions that bundle pre-configured model providers, private agents, and governed tool connectors without modifying the core open-source codebase.

Because Berd functions as an endpoint application rather than a cloud-hosted service, organisations maintain full oversight regarding which external AI models are authorised for data transmission.

From Desktop Workspaces to Multiplayer Collaboration

While Berd optimises the single-player desktop experience, Block is simultaneously advancing Buzz, an open-source collaboration platform built on the decentralized Nostr protocol.

Where Berd provides high-privilege access to a single user’s local environment, Buzz enables humans and cryptographic AI agents to collaborate in shared multiplayer rooms. Block plans to integrate Berd’s desktop capabilities into the Buzz ecosystem over time, allowing individual work to move seamlessly into shared team workflows.

By decoupling the agent interface from specific model vendors, Berd presents a flexible approach to managing digital AI assistants directly from the desktop.

Disclaimer: This article is provided for informational purposes only, mistakes may be made, and it's not offered or intended to be used as legal, tax, investment, financial, or any other advice.