Block open-sources Berd, a local-first desktop workspace for multi-harness AI agents

Block just open-sourced Berd, a desktop application that gives you a single workspace for running AI agents across different models and harnesses. Unlike vendor-locked tools, Berd is free, stores conversation history locally, and lets you bring your own providers. For builders juggling Goose, Claude Code, and Codex, this is the first open-source layer that sits above the agent harness rather than replacing it.

A desktop workspace that separates the environment from the harness

Berd is built with Tauri 2 and React 19 and communicates with Goose through the Agent Client Protocol (ACP). Goose runs as a sidecar and handles the agent loop via the Model Context Protocol (MCP), supporting both proprietary and local models. Berd itself manages projects, sessions, context, agents, and configuration. The repository reached version 0.6.2 on August 18 with 91 contributors, and builds are available for macOS, Windows, and Linux under Apache 2.0.

The design emphasizes visible operational state: you can always see which project, files, agent, model, and provider are active. Configuration is treated as part of the workflow, not buried in settings. Berd also gives agents distinct visual identities (called Gloopies) with roles like Pushback (devil's advocate) or Choosey (decision helper), making it easier to tell them apart when running multiple agents.

Why this matters for AI builders

The key differentiator is that Berd is not tied to a single vendor's models or harness. You can use Goose, Claude Code, Codex, or any other ACP-compatible harness, and configure any model provider. This contrasts with OpenAI's Codex app (OpenAI models only), Anthropic's Claude Code (Claude only), or Cursor (subscription-based). Berd is free and open source; you only pay your model provider costs.

For teams that need to switch between models depending on the task, or that want to keep sensitive data on-device, Berd's local-first architecture is practical. Conversation history stays in the local Goose session database, credentials live in the OS keychain, and telemetry is off by default. Block says telemetry excludes prompts, files, source code, and personal identifiers.

Practical implications and caveats

Berd is not a complete enterprise management platform. There is no SSO, centralized policy enforcement, or audit logging described. Block provides distribution seams for organizations to overlay managed settings, but central governance after deployment is not documented. The repository does not accept pull requests; Block's small team controls the upstream. For long-term infrastructure decisions, that matters.

Block's roadmap points toward integrating Berd's single-player experience into Buzz, its multiplayer collaboration platform. Berd remains available as a desktop app, but Buzz is the product Block encourages for new users. If you need a local, multi-harness workspace today, Berd is worth evaluating. If you need team collaboration, wait for Buzz integration.

Also note: Berd does not make AI processing local. When you send a prompt to a configured model, data leaves your device according to that provider's terms. Local-first means storage, not computation.