Block (formerly Square) has open-sourced Berd, a local-first, Apache 2.0-licensed desktop application designed as a universal "daily work surface" for AI agents. Rather than forcing users into a single AI ecosystem or browser tab, Berd acts as an orchestration layer that unifies agents like Goose, Claude Code, and Codex across multiple foundation models. It prioritizes data privacy by keeping session history and credentials on your local machine, while bringing distinct visual personalities for your AI Agents (like animated "Gloopies") to make them instandly recognizable and manageable. This may be a way to assemble all your individual tools and AI agents with different capabilities under one roof. Use cases might be: Video creation, voice overs, content creation, social media posting all in a single harness running on your own machine. Key Features & Benefits Model & Harness Agnostic: Works seamlessly across proprietary and open-source models, letting you switch between Anthropic, OpenAI, or local runtimes using standard protocols (ACP/MCP). Local-First Privacy: Stores session history, conversation logs, and credentials locally on your computer (using the OS keychain), ensuring telemetry is off by default and your data isn't harvested by default. <--- don't lose your sessions over time. Persistent Workspaces: Organize files, tools, repositories, and context into persistent projects so you don't have to rebuild context for every new prompt. Visual Agent Personas: Assign distinct roles, skills, and visual character identities (such as "Gloopies") to agents so you instantly know which agent handles which task (e.g., code review, decision-narrowing, or writing style). It comes with 7 agents I understand...still experimenting! Enterprise Customization ("Distribution Seams"): Open-source (Apache 2.0) for macOS, Windows, and Linux—allowing companies to package custom, governed distributions with pre-configured models and tool permissions. Links Read the full VentureBeat article: Block's new Apache 2.0 agent workspace Berd works across models...