Block, the Jack Dorsey company behind Square and Cash App, has open-sourced Berd, a desktop AI workspace built internally to unify agent work across models and harnesses. It is available now on GitHub under Apache 2.0, with signed builds for macOS, Windows and Linux. Version 0.6.2 shipped August 18, the seventh public release, with 91 listed contributors. The stack is Tauri 2 and React 19. It bundles Goose, Block's own agent framework, as a pinned sidecar process, with the backend version locked and hash-verified at build time so the agent layer cannot silently drift.
The design rationale is worth reading in detail because it is more specific than most enterprise AI product pitches. Berd explicitly surfaces provider state, session state, model selection and failure conditions rather than hiding them behind assistant-style polish. Projects are persistent: files, instructions and agent configurations survive between sessions without manual reconstruction. Agents get visual identities, roles, skills and tools rather than appearing as identical chat windows. The product spec names what it is refusing to be: a toy, a chatbot wrapper, a mobile-first product. Block's Head of AI Capabilities Brad Axen draws a hard line between agent harnesses like Goose, Claude Code and Codex, which manage the model loop, and Berd, which provides the consistent desktop surface around them.
The open-source release means organizations can inspect exactly what runs locally, substitute their own Goose binary for testing and configure any supported provider or harness rather than accepting a Block-selected model stack. Block has also contributed Goose to the Agentic AI Foundation, the Linux Foundation body it helped establish in December 2025 alongside Anthropic and OpenAI. The full article includes the product specification language, Block's internal account of how fragmented multi-agent workflows drove the build decision and a direct comparison against Codex, Claude Code and Cursor. That comparison is the section to read if you are evaluating where Berd actually fits.
[READ ORIGINAL →]